Reasoning about Rational Agents

Intelligent Robots and Autonomous Agents

Ronald C. Arkin, editor

Behavior-Based Robotics, Ronald C. Arkin, 1998

Robot Shaping: An Experiment in Behavior Engineering, Marco Dorigo and

Marco Colombetti, 1998

Layered Learning in Multiagent Systems: A Winning Approach to Robotic

Soccer, Peter Stone, 2000

Evolutionary Robotics: T he Biology, Intelligence, and Technology of Self­

Organizing Machines, Stefano Nolfi and Dario Floreano, 2000

Reasoning about Rational Agents, Michael Wooldridge, 2000

Reasoning about Rational Agents

Michael Wooldridge

The MIT Press

Cambridge, Massachusetts
London, England
@ 2000 Massachusetts Institute of Technology

All rights reserved. No part of this book may be reproduced in any form by any electronic
or mechanical means (including photocopying, recording, or information storage and retrieval)
without permission in writing from the publisher.

This book was set in Times Roman by the author using the IHEX document preparation system
and printed and bound in the United States of America.

Library of Congress Cataloging-in-Publication Data

Wooldridge, Michael 1., 1966-

Reasoning about rational agents / Michael Wooldridge.
p. cm.---(Intelligent robots and autonomous agents)
Includes bibliographical references and index.
ISBN 0-262-23213-8 (hc: alk. paper)
1. Intelligent agents (Computer software). I. Title. II. Series.
QA76.76I58W66 2000
006.3-dc21 00-035474
CIP
To Janine
Contents

Preface IX

Acknowledgments Xlll

1 Rational Agents

2 The Belief-Desire-Intention Model 21

3 Introduction to £ORA 47

4 £ORA Defined 69

5 Properties of Rational Agents 91

6 Collective Mental States 1 13

7 Communication 125

8 Cooperation 147

9 Logic and Agent Theory 163

Appendix A:
Summary of Notation 179

Appendix B:
Formal Foundations 18 1

References 205
Index 223
Preface

This book is about a certain class of computer systems that are known as
agents, and more specifically, rational agents. We call these computer systems
agents because they are capable of independent, autonomous action in order
to meet their design objectives: put crudely, they are capable of deciding for
themselves what to do in any given situation. We call them rational agents
because they make good decisions about what to do.
Engineering rational agents may at first sight appear to be a somewhat
abstract, even fanciful endeavor, divorced from the everyday reality of com­
puter science and software engineering. This is not the case. Rational agents
are finding a wide range of applications. A famous example is NASA's Deep
Space 1 (OS 1) mission. Launched from Cape Canaveral on 24 October 1998,
OS 1 was the first space probe to have an autonomous, agent-based control sys­
tem [ 165]. Before osl, space missions required a ground crew of up to 3 00
staff to continually monitor progress. This ground crew made all necessary
control decisions on behalf of the probe, and painstakingly transmitted these
decisions to the probe for subsequent execution. Given the length of typical
planetary exploration missions, such a procedure was expensive and, if the de­
cisions were ever required quickly, it was simply not practical. The autonomous
control system in os 1 was capable of making many important decisions itself.
This made the mission more robust, particularly against sudden unexpected
problems, and also had the very desirable side effect of reducing overall mis­
sion costs. The os 1 example may seem somewhat extreme: after all, spacecraft
control systems are hardly an everyday software engineering problem. But the
same technology is finding applications in much more commonplace domains.
For example, such agents are increasingly being used for Internet-based elec­
tronic commerce applications, where they autonomously buy and sell goods
on behalf of a user [ 168].
As it turns out, for all but the simplest of applications, it is very diffi­
cult to build computer systems that can autonomously make good decisions
about what to do. For sure, there are a number of very powerful mathemati­
cal frameworks available, which can help us to understand and analyze many
decision-making scenarios. Decision theory is an obvious example. Decision
theory tells us that a rational agent is one that chooses an action which maxi­
mizes expected utility, where expected utility is defined in terms of the actions
available to the agent, the probability of certain outcomes, and the preferences
the agent has with respect to these outcomes. In multiagent scenarios, where
an agent is required to interact with other agents, game theory has proved to
be a powerful predictive and analytic tool. However, valuable and important
x Preface

though tools such as decision theory and game theory are, they give us few
clues with respect to implementing rational agents.
This book focuses on a model of rational decision making as practical
reasoning, of the kind that humans engage in every day. The model is called
the belief-desire-intention (8DI) model, because it acknowledges the primacy
of beliefs, desires, and, crucially, intentions, in the practical reasoning process.
The 8DI model is particularly interesting because it combines three distinct
components:

• A philosophical component.
The 8DI model is based on a widely respected theory of rational action in
humans, developed by the philosopher Michael Bratman [20].

• A software architecture component.

The 8DI model of agency does not prescribe a specific implementation. The
model may be realized in many different ways, and indeed a number of
different implementations of it have been developed. However, the fact that
the 8DI model has been implemented successfully is a significant point in its
favor. Moreover, the 8DI model has been used to build a number of significant
real-world applications, including such demanding problems as fault diagnosis
on the space shuttle.

• A logical component.
The third component of the 8DI model is a family of logics. These logics
capture the key aspects of the 8DI model as a set of logical axioms. There
are many candidates for a formal theory of rational agency, but 8DI logics in
various forms have proved to be among the most useful, longest-lived, and
widely accepted.

The emphasis in this book is primarily on the logical foundations of 8DI

systems. 8DI logics were developed by Anand Rao and Michael Georgeff,
in order to give an abstract, idealized semantics to the 8DI agents they were
building throughout the early 1990s at the Australian AI Institute [80]. Building
on the work of researchers in both theoretical computer science and artificial
intelligence, they were able to demonstrate how a particular type of multi­
modal logic could be used to capture many aspects of Bratman's 8DI theory.
In this book, I build upon and significantly extend the basic logical frame­
work of Rao and Georgeff. I begin by introducing the concept of an agent, and
give some basic background both to the philosophical foundations and soft­
ware design of 8DI agents. After introducing the basic 8DI model, I present
Preface xi

CORA, the formal framework that will be used throughout the remainder of
the book. CORA - which stands for "Cogic Of Rational Agents" - is a 8DI
logic, which extends Rao and Georgeff's original formalism. CORA allows
us to represent and reason about the beliefs, desires, intentions, and actions of
agents within a system, and how these beliefs, desires, intentions, and actions
change over time.
Once the basic logical framework is introduced, the remainder of the book
is taken up largely with an investigation of how it can be used to capture a range
of important aspects of rational agency. These aspects can be divided into those
that relate to individual agents and those that relate to multiple agents. With
respect to individual agents, I show how CORA can be used to capture many
of our intuitions about the relationships among beliefs, desires, and intentions,
as well as other concepts such as ability. The major contribution of the book
comes from the focus on multiagent aspects. Specifically, I show how CORA
can be used to capture:

• mutual mental states and the closely related notion of teamwork;

• communication between agents, in the form of speech acts; and
• cooperative problem solving, whereby a group of agents collaborate to
achieve some task.

There is some tension between the desire to write a book that is rigorous from
the perspective of mathematical logic, and the desire to write a book that is
readable. Logical rigor all too often seems to preclude readability: the level of
detail required to pass muster with logicians can be daunting to the uninitiated.
In the event, I have done my best to compromise. Thus, as well as giving formal
definitions and complete proofs of all results, I have tried wherever possible
to explain at length the intuition behind the formalism, and to give extensive
pointers to further reading. Those sections that contain a significant proportion
of abstract mathematical material are separated out from the main text, and
indicated in the relevant section header by means of an asterisk ("*").
8DI is probably the most-implemented framework for building rational
agents. I hope that after reading this book, you will understand why so many
researchers believe it represents a powerful framework through which to un­
derstand, implement, and reason about rational agency.
Acknowledgments

First, thanks to my colleagues from London, Manchester, Liverpool, and else­

where, who have provided me with a friendly and supportive environment in
which to work over the years. In particular, I would like to thank Innes Fergu­
son, Michael Fisher, Afsaneh Haddadi, Nick Jennings, Alessio Lomuscio, Jorg
Miiller, and Simon Parsons, all of whom I have had the pleasure of working
with closely.
Mike Georgeff and Anand Rao have both been generous in their encour­
agement of my research over the years. As well as owing an obvious intellec­
tual debt to them, this book has also therefore benefited from their experience
in less obvious, but no less important ways.
Many other colleagues from the international academic community have
helped to shape my ideas both about multiagent systems and about the aca­
demic profession in general. I should at least thank Cristiano Castelfranchi,
Keith Decker, Ed Durfee, Dov Gabbay, Wiebe van der Hoek, Michael Huhns,
Mark d'lnverno, Sarit Kraus, Yves Lesperance, Mike Luck, John-Jules Meyer,
Carles Sierra, Munindar Singh, Katia Sycara, Milind Tambe, and Gerhard
WeiB. (If! missed anyone . . . you know who you are!)
I started work on this book while on a short sabbatical at the University
of Melbourne in Australia. I would like to thank Leon Sterling for the oppor­
tunity to visit his Department, and for the warm reception and hospitality I
received throughout my stay. Thanks also to Liz Sonenberg and David Kinny
for their support - David in particular for showing me around and making me
welcome.
When I first conceived this book, I was employed at Queen Mary and
Westfield College, University of London; thanks to David Olver and Laurie
Cuthbert for their support while at Q MW.
Bob Prior at MIT Press has been a source of encouragement and good
humor throughout the writing of this book, for which I am very grateful. In
addition to Bob, I would like to thank Katherine Innis and Deborah Cantor­
Adams from MITP for their help and feedback.
Thanks to all those who read and commented on drafts of this book,
sometimes when those drafts should perhaps not have seen the light of day:
Trevor Bench-Capon, Rogier van Eijk, Peter McBurney, Simon Parsons, and
Martijn Schut.
Finally, to my wife Janine, who means everything to me: thank you.

Michael Wooldridge
Liverpool, Spring 2000
Reasoning about Rational Agents
1 Rational Agents

The world we inhabit is full of agents. We encounter them every day, as we

go about our lives. By the term "agent," I mean an entity that acts upon the
environment it inhabits. Agents are not merely observers of their environment,
nor are they the passive recipients of actions performed by other entities.
Rather, agents are the active, purposeful originators of action. These actions
are performed in order to modify and shape the environment inhabited by the
agent. People like you and I are the most obvious examples of agents in the
real world, but there are many others, such as legal entities like governments,
companies, and corporations.
In our dealings with the world, we often make a distinction between agents
that are rational and those that are not, in the following sense. An agent is said
to be rational if it chooses to perform actions that are in its own best interests,
given the beliefs it has about the world. For example, if! have a goal of staying
dry, and I believe it is raining, then it is rational of me to take an umbrella
when I leave my house. Taking the umbrella will enable me to satisfy my goal
of staying dry, and in this sense, it is in my best interest. You would still be
inclined to refer to my behavior as rational even if I was mistaken in my belief
that it was raining: the point is that I made a decision that, if my beliefs were
correct, would have achieved one of my goals.
In one sense, human agents are notoriously irrational when they make
decisions. From time to time, we all make decisions that, with the benefit
of hindsight, are self-evidently poor. When evaluated against strict metrics of
rationality, humans rarely perform well. But this paints a misleading picture.
The fact is that, overall, humans are very good decision makers. We inhabit a
continuously changing physical world of extraordinary richness, complexity,
and diversity. Our plans are constantly thwarted, both by random quirks of
nature and, occasionally, by the deliberately malicious intervention of our
peers. We are forced to make decisions for which we are ill prepared, both
in terms of prior experience and in terms of the information upon which to
base decisions. And yet, despite all this, we prosper. As decision makers, we
seem to be quite robust to the vagaries of our environment. We successfully
interact with our peers even when they are in conflict with us, and we cope
with uncertainty about the world, failed actions and plans, and the complexity
of our environment.
2 Chapter 1

One goal of the artificial intelligence (AI) community is to engineer com­

puter programs that can act as autonomous, rational agents. We wish to build
computer programs that can independently make good decisions about what
actions to perform. But it is not enough to have programs that think of a good
action to perform - we wish to have them actually execute these actions. In
other words, we want to create rational agents that are embodied in some envi­
ronment - that inhabit and act upon some environment in the same way that
we inhabit and act upon ours.
Ideally, we would like to engineer agents that are as good at making
decisions and acting upon them as we are. Of course, this goal is a long way
from being achieved. Unfortunately, like so many other problems encountered
throughout the history of AI, it has proven extraordinarily difficult to engineer
agents that can select and execute good decisions for moderately complex
environments. Nor is the problem likely to be solved in any meaningful sense
for some time yet.

1.1 Properties of Rational Agents

In order to understand why it is hard to engineer a rational agent, let us

consider what sorts of properties we expect a rational agent to exhibit. First, it
is important to understand that we usually consider agents to be systems that
are situated or embodied in some environment - agents are not disembodied
systems. By this, I mean that agents are capable of sensing their environment
and have a repertoire of possible actions that they can perform in order to
modify the environment. This leads to the view of an agent as shown in
Figure 1.1. The requirement that agents must be embodied implies that many
of the systems that have been developed to date within AI do not count as
agents. Examples include most theorem provers and expert systems.
The environment that an agent occupies may be physical (in the case of
robots inhabiting the physical world) or a software environment (in the case of
a software agent inhabiting a computer operating system or network). In the
case of a physically embodied agent, actions will be physical, such as moving
objects around. The sensor input received by the agent will be comprised
of video feeds and the like. In the case of a software agent, actions will be
software commands (such as the UNIX rm command, which removes a file),
and sensor input will be obtained by performing commands such as 15 (which
obtains a directory listing) [53].
Rational Agents 3

AGENT

action
output

ENVIRONMENT

Figure 1.1
Agents and their environments. The agent takes sensory input from the environment, and
produces as output actions that affect it.

Note that in almost all realistic applications, agents have at best partial
control over their environment. Thus, while they can perform actions that
change their environment, they cannot in general completely control it. It
therefore makes sense to divide the environment, conceptually at least, into
those parts that the agent controls, those parts it partially controls, and those
parts over which it has no control. (If we regard an agent's internal state as
being part of the environment, then this would be controlled by the agent.)
Apart from being situated in an environment, what other properties do we
expect a rational agent to have? Wooldridge and Jennings [249] argue that
agents should have the following properties:

• autonomy;
• proactiveness;
• reactivity; and
• social ability.

At its simplest, autonomy means nothing more than being able to operate in­
dependently. Thus, at the very least, an autonomous agent makes independent
decisions - its decisions (and hence its actions) are under its own control,
and are not driven by others. Autonomous agents are therefore the loci of de­
cision making. However, autonomy is usually taken to have a slightly stronger
meaning than this. We generally assume that an autonomous agent is one that
has its own beliefs, desires, and intentions, which are not subservient to those
4 Chapter 1

of other agents. This is not to say that these beliefs, desires, and intentions are
necessarily represented within an agent - the point is that we regard a rational
agent as "having its own agenda," which may or may not concur with that of
other agents.
Proactiveness means being able to exhibit goal-directed behavior. If an
agent has a particular goal or intention, then we expect the agent to try to
achieve this intention. For example, if I have an intention to write a book,
then it is rational of me to actually act towards this intention, by eventually
putting pen to paper. Moreover, proactiveness means exploiting serendipity.
For example, if I have a desire to become rich, and an opportunity to make a
fortune arises, then it is only rational of me to try to exploit this opportunity,
perhaps by generating a new goal. Proactiveness rules out entirely passive
agents, who never try to do anything.
Being reactive means being responsive to changes in the environment.
As I noted above, in everyday life, our plans rarely run smoothly. They are
frequently thwarted, both by the vagaries of nature and by other agents fouling
them up. W hen we become aware that our plans have gone awry, we do not
ignore this fact. We respond, by choosing an alternative course of action.
Responses may be of the "stimulus ---+ response" style, or they may involve
further deliberation. For example, if I am driving my car and I realize I am
about to crash into something, I will swerve to avoid it. It is unlikely that this
kind of response involves much deliberation on my part; the behavior seems
to be "hardwired." In contrast, if! am traveling to Heathrow Airport for a 9.30
AM flight, and I discover that the train I was about to catch is not running, I
will actively deliberate in order to fix upon a course of action that will still
enable me to catch my flight. This deliberation typically involves at least some
reasoning about the alternatives available (catching a taxi or bus, for example).
Crucially, however, I fix upon an alternative sufficiently quickly that it will
deliver me to the airport in time - I will not deliberate for so long that it
becomes impossible to catch the flight.
Designing a purely reactive system, which simply responds to environ­
mental stimuli, is not hard. We can implement such a system as a lookup table,
which simply maps environment states directly to actions. Similarly, devel­
oping a purely goal-driven system is not hard. (After all, this is ultimately
what conventional computer programs are.) However, implementing a system
that achieves an effective balance between goal-directed and reactive behavior
turns out to be very hard. We elaborate on this problem in the next chapter.
Rational Agents 5

Finally, let us say something about social ability. In one sense, social
ability is trivial. After all, every day, millions of computers across the world
routinely exchange information with both humans and other computers. But
the ability to exchange bit streams is not really social ability. Consider that,
in the human world, comparatively few of our goals can be achieved without
interaction with other people and organizations, who cannot be assumed to
share our goals. These agents are themselves autonomous, with their own
agendas to pursue. To achieve our goals in such situations, we must negotiate
and cooperate with these agents. We may be required to understand and reason
about their beliefs and goals, and to perform actions (such as paying them
money) that we would not otherwise choose to perform, in order to get them
to cooperate with us, and achieve our goals. This type of social ability -
the ability to interact (through negotiation and cooperation) with other self­
interested agents - is much more complex and much less understood than
simply the ability to exchange bit streams. In the later chapters of this book, I
focus on such interactions in detail.

1.2 A Software Engineering Perspective

It is important to understand that the engineering of rational agents is not

simply an abstract intellectual pursuit. It arises as a result of problems faced
by software engineers across the world at the start of the new millennium.
Originally, software engineering concerned itself primarily with what are
known as "functional" systems. A functional system is one that simply takes
some input, performs some computation over this input, and eventually pro­
duces some output. Such systems may formally be viewed as functions f :
I ---+ 0 from a set I of inputs to a set 0 of outputs. The classic example of
such a system is a compiler, which can be viewed as a mapping from a set I of
legal source programs to a set 0 of corresponding object or machine code pro­
grams. Although the internal complexity of a functional system may be great
(e.g., in the case of a compiler for a complex programming language such as
ADA), functional programs are, in general, comparatively simple to engineer
correctly and efficiently. Unfortunately, many of the computer systems that
we desire to build are not functional in the sense we describe here. Rather than
simply computing a function of some input and then terminating, such systems
are reactive, in the following sense:!

1 There are at least three current usages of the term "reactive system" in computer science. The
6 Chapter 1

Reactive systems are systems that cannot adequately be described by the relational or
functional view. The relational view regards programs as functions ... from an initial
state to a terminal state. Typically, the main role of reactive systems is to maintain an
interaction with their environment, and therefore must be described (and specified) in
terms of their on-going behavior ... [E]very concurrent system ... must be studied by
behavioral means. This is because each individual module in a concurrent system is a
reactive subsystem, interacting with its own environment which consists of the other
modules. [ 177]

Pnuelian reactive systems have long been recognized to be harder to engineer

correctly and efficiently than functional systems. Indeed, Pnuelian reactive
systems form a major research theme in their own right within mainstream
computer science [11, 144, 145].
It should be clear that all agents (in the sense that I describe them) are
Pnuelian reactive systems, although of course not all Pnuelian reactive systems
are agents. Researchers and developers throughout many subfields of computer
science are beginning to recognize that agents represent an important new way
of thinking about and designing Pnuelian reactive systems in particular, and
distributed computer systems in general. There are several reasons for this.
First, we delegate increasingly sophisticated and complex activities to com­
puters, where we require them to act independently on our behalf. Agents are a
very natural metaphor for thinking about such systems. Second, agents reflect
an important ongoing trend in computer science toward the understanding that
interacting systems are the norm, rather than the exception. The rapid growth
of the Internet is the most obvious public sign of the change in emphasis to­
wards interacting, distributed systems. Interaction is a fundamental research
issue in the agent community. Of particular interest are systems that can co­
operate and negotiate with each other. As the Internet continues to grow in
importance over the coming decades, the ability of computer systems to au­
tonomously cooperate and negotiate will provide enormous opportunities in
such areas as electronic commerce [168].

first, oldest usage is that by Pnueli and followers (see, e. g., [177], and the description above).
Second, researchers in AI planning take a reactive system to be one that is capable of responding
rapidly to changes in its environment - here the word "reactive" is taken to be synonymous
with "responsive" (see, e. g., [l12]). More recently, the term has been used to denote systems that
respond directly to their environment, rather than reason explicitly about it (see, e. g., [40]). I use
the adjective "Pnuelian" when referring to reactive systems in the sense that Pnueli describes.
Rational Agents 7

1.3 Belief-Desire-Intention Agents

In this book, I focus on the belief-desire-intention (8DI) model of rational

agency [22]. The 8DI model gets its name from the fact that it recognizes the
primacy of beliefs, desires, and intentions in rational action.
Intuitively, an agent's beliefs correspond to information the agent has about
the world. These beliefs may be incomplete or incorrect. An agent's desires
represent states of affairs that the agent would, in an ideal world, wish to
be brought about. (Implemented 8DI agents require that desires be consistent
with one another, although human desires often fail in this respect.) Finally,
an agent's intentions represent desires that it has committed to achieving. The
intuition is that an agent will not, in general, be able to achieve all its desires,
even if these desires are consistent. Ultimately, an agent must therefore fix
upon some subset of its desires and commit resources to achieving them. These
chosen desires, to which the agent has some commitment, are intentions [35].
What makes the 8DI model particularly compelling is that it combines
three important elements:

1. It is founded upon a well-known and highly respected theory of rational

action in humans.
2. It has been implemented and successfully used in a number of complex
fielded applications.
3. The theory has been rigorously formalized in a family of 8DI logics.

The 8DI theory of human rational action was originally developed by Michael
Bratman [20]. It is a theory of practical reasoning - the process of reasoning
that we all go through in our everyday lives, deciding moment by moment
which action to perform next. Bratman's theory focuses in particular on the
role that intentions play in practical reasoning. Bratman argues that intentions
are important because they constrain the reasoning an agent is required to do in
order to select an action to perform. For example, suppose I have an intention
to write a book. Then while deciding what to do, I need not expend any effort
considering actions that are incompatible with this intention (such as having
a summer holiday, or enjoying a social life). This reduction in the number
of possibilities I have to consider makes my decision making considerably
simpler than would otherwise be the case. Since any real agent we might care to
consider - and in particular, any agent that we can implement on a computer
- must have resource bounds, an intention-based model of agency, which
8 Chapter 1

constrains decision-making in the manner described, seems attractive.

The BDI model has been implemented several times. Originally, it was re­
alized in IRMA, the Intelligent Resource-bounded Machine Architecture [22].
IRMA was intended as a more or less direct realization of Bratman's theory of
practical reasoning. However, the best-known implementation is the Procedu­
ral Reasoning System (PRS) [78] and its many descendants [61, 183, 46, 100].
In the PRS, an agent has data structures that explicitly correspond to beliefs,
desires, and intentions. A PRS agent's beliefs are directly represented in the
form of PROLOG-like facts [34, p.3]. Desires and intentions in PRS are real­
ized through the use of a plan library. 2 A plan library, as its name suggests,
is a collection of plans. Each plan is a recipe that can be used by the agent to
achieve some particular state of affairs. A plan in the PRS is characterized by
a body and an invocation condition. The body of a plan is a course of action
that can be used by the agent to achieve some particular state of affairs. The
invocation condition of a plan defines the circumstances under which the agent
should "consider" the plan. Control in the PRS proceeds by the agent contin­
ually updating its internal beliefs, and then looking to see which plans have
invocation conditions that correspond to these beliefs. The set of plans made
active in this way correspond to the desires of the agent. Each desire defines a
possible course of action that the agent may follow. On each control cycle, the
PRS picks one of these desires, and pushes it onto an execution stack, for sub­
sequent execution. The execution stack contains desires that have been chosen
by the agent, and thus corresponds to the agent's intentions.
The third and final component of the BDI model is the logical component.
This logical component provides a family of tools that allow us to reason about
BDI agents.

1.4 Reasoning about Belief-Desire-Intention Agents

Computer science is, as much as it is about anything, about developing theories

and formalisms that allow us to reason about the behavior of computational
systems. The theory of a system can be understood as the semantics of that
system. An obvious question, therefore, is: What sort of theory can we use to
give a semantics to BDI systems? The answer presented by this book is to use
BDI logics to axiomatize properties of agents.

2 In this description of the PRS, I have modified the original terminology somewhat, to be more
in line with contemporary usage; I have also simplified the control cycle of the PRS slightly.
Rational Agents 9

A fundamental problem in developing such a theory of rational agency is to

give an account of the relationships that exist between an agent's mental states.
In particular, a complete agent theory would explain how an agent's mental
states lead it to select and perform rational actions, thus realizing the mapping
from perception to action illustrated in Figure 1.1. To give a flavor of the issues
involved in developing such a theory, consider the relationship between belief
and intention. Suppose I intend to bring about some state of affairs <po What
does the fact that I intend <p imply about my beliefs with respect to <p? Well,
clearly, it would make no sense for me to intend <p if I believe that <p is already
true. If my book is written, then there is no point in me intending to write it.
This property - that intending <p implies you do not already believe <p - is
a constraint that should be satisfied by a rational agent. We can capture this
constraint in a formula schema of .cORA, the logic I use throughout this book
to represent the properties of agents.

(Int i <p) => ...,(Bel i <p) (Ll)

Here, (Int i <p) is a construction that means "agent i intends <p." It is not hard
to guess that (Bel i <p) means "agent i believes <p." A "complete" theory of
rational agency would be comprised of a number of formula schemas such as
(1.1), which would together axiomatize the properties of rational agency.
Unfortunately, giving anything like a complete account of the relationships
between an agent's mental states is extremely difficult. Part of the problem
is that the objects of study - beliefs, desires, and the like - are rather
different to phenomena that we observe in the physical world. In attempting
to develop formal theories of such notions, we are forced to rely very much on
our intuitions about them. As a consequence, such theories are hard to validate
(or invalidate) in the way that good scientific theories should be. Fortunately,
we have powerful tools available to help in our investigation. Mathematical
logic allows us to represent our theories in a transparent, readable form, and
examine (through the medium of formal proof) the predictions that our theories
make, in order to see whether or not their consequences make sense.
.cORA stands for ".cogic Of Rational Agents." .cORA is a 8DI logic:
originally developed by Anand Rao and Michael Georgeff [187, 186, 188,
189, 184, 185], 8DI logics allow us to represent the beliefs, desires, and
intentions of agents. In addition to the 8DI component, .cORA contains a
temporal component, which allows us to represent the dynamics of agents and
their environments - how they change over time. Finally, .cORA contains
an action component, which allows us to represent the actions that agents
10 Chapter 1

perform, and the effects that these actions have.

1.5 Frequently Asked Questions (FAQ)

Certain questions are frequently asked about the work presented in this book,
and so it seems only appropriate that I should include a FAQ.

Why Not Decision Theory?

Decision theory is a mathematical theory of rational decision making. Decision

theory defines a rational agent as one that maximizes expected utility. The basic
ideas of decision theory are very simple. Assume we have some agent, and that
Ac is the set of all possible actions available to it. The performance of an action
by the agent may result in a number of possible outcomes, where the set of all
outcomes is n = {w, w', ...}. Let the probability of outcome wEn given that
the agent performs action a E A c be denoted by P(w I a), and finally, let the
utility of an outcome wEn for the agent be given by a function U : n -+ JR.
If U(w) > U(w'), then the agent prefers outcome w over outcome w'.
The expected utility of an action a E A c is denoted EU(a). Thus EU(a)
represents the utility that an agent could expect to obtain by performing action
a.

EU(a) = L U(w)P(w I a) (1.2)

wEO

A rational agent is then one that chooses to perform an action a that maximizes
EU( ...). It is straightforward to define the behavior of such an agent. Let
the function !apt take as input a set of possible actions, a set of outcomes,
a probability distribution over possible outcomes given the performance of
actions, and a utility function, and let this function return an action. Then we
can define the desired behavior of !apt as follows.

!opt(Ac,n,p, U) = argmax
aEAc
L U(w)P(w I a) (1.3)
wEO

As Russell and Norvig point out, equations (1.2) and (1.3) could be seen in one
sense as defining the whole of artificial intelligence [198, pA72]. It seems that
in order to build a rational agent, all we need to do is implement the function
!opt.But while decision theory is a normative theory of rational action (it tells us
what an agent should in principle do) it has nothing to say about how we might
efficiently implement the function!opt. The problem is that!apt seems to require
Rational Agents 11

an unconstrained search over the space of all actions and their outcomes. Such
a search is prohibitively expensive, particularly in the case where an agent
needs to consider sequences of actions (i.e., plans).
Another problem is that it is hard in practice to obtain the probability
distribution P or utility function U. If we consider a domain defined by m
attributes, each of which may take n values, then the set n will contain
n
m different outcomes. Simply enumerating these outcomes is likely to be
impractical in many circumstances. Even if it is possible to enumerate the
outcomes, actually obtaining the required probabilities and utilities for them
is notoriously hard [198, pp.475--480]. See the "Notes and Further Reading"
section at the end of this chapter for references to how decision theory is
currently being used in artificial intelligence.

Why Not Game Theory?

Game theory is a close relative of decision theory, which studies interactions

between agents where each agent is accorded a utility function as described
above [14]. The tools and techniques of game theory have found many applica­
tions in computational multiagent systems research, particularly when applied
to such problems as negotiation [195, 199]. Game theory shares with decision
theory many concepts, in particular the concept of a rational agent as one that
acts to maximize expected utility. However, game theory also shares many of
the problems of decision theory. While it tells us what a rational agent should
do in principle, it does not give us many clues as to how to implement such
decision makers efficiently. In addition, like decision theory, game theory is
quantitative in nature.
It is important to note that all this should not be taken to imply that game
theory is in any sense wrong, or that it is of no use. On the contrary, game
theory is an extremely powerful analytic tool, which will certainly be applied
more widely in years to come. But it does have limitations.

Why Logic?

There are some in the AI research community who believe that logic is (to put
it crudely) the work of the devil, and that the effort devoted to such problems
as logical knowledge representation and theorem proving over the years has
been, at best, a waste of time. At least a brief justification for the use of logic
therefore seems necessary.
12 Chapter 1

First, by fixing on a structured, well-defined artificial language (as opposed

to unstructured, ill-defined natural language), it is possible to investigate the
question of what can be expressed in a rigorous, mathematical way (see, for
example, Emerson and Halpern [50], where the expressive power of a number
of temporal logics are compared formally). Another major advantage is that
any ambiguity can be removed (see, e.g., proofs of the unique readability of
propositional logic and first-order predicate logic [52, pp.39--43]).
Transparency is another advantage:
By expressing the properties of agents, and multiagent systems as logical axioms
and theorems in a language with clear semantics, the focal points of [the theory] are
explicit. The theory is transparent; properties, interrelationships, and inferences are
open to examination. This contrasts with the use of computer code, which requires
implementational and control aspects within which the issues to be tested can often
become confused. [68, p.88]

Finally, by adopting a logic-based approach, one makes available all the results
and techniques of what is arguably the oldest, richest, most fundamental, and
best-established branch of mathematics.
The final chapter in the main text of this book, chapter 9, considers in detail
the role of logic in the theory and practice of multiagent systems.

Why Not First-Order Logic?

Those active in the area of logic for AI can be divided broadly into two
camps: those who make use of classical (usually first-order) logic, and those
who use non-classical logics (in particular, modal and temporal logics of the
kind used in this book) of some sort. This book makes use of a logic called
.cORA. From a technical point of view, .cORA is a first-order branching time
logic, containing modal connectives for representing the beliefs, desires, and
intentions of agents, as well as a dynamic logic-style apparatus for representing
and reasoning about the actions that agents perform.
Strong opinions are held in both camps on the relative merits of classical
versus non-classical logics. Those in favor of classical logic point to a number
of general arguments against modal logics:

1. First-order logic is sufficiently expressive that it can be used to encode

almost any form of knowledge required.
2. While the satisfiability problem for first-order logic is only semi-decidable
(see, e.g., [73, p.58]) the corresponding problem for modal and temporal logics
tends to be even worse. Some multimodal logics are undecidable even in the
Rational Agents 13

propositional case [91]. This makes it doubtful that there will ever be practical
theorem-proving methods for such logics.

There is certainly some strength to the first argument. With sufficient ingenuity,
one can use first-order logic to encode anything that can be encoded using
a modal or temporal logic. In this sense, first-order logic is a truly generic
formalism. But this argument misses the point slightly. To see what I mean,
consider the following English-language statement.

We are not friends until you apologize. (1.4)

We can encode (1.4) in first-order logic by translating it to the following

statement: there is some time t such that you apologize at t and t is later than
now, such that for all time points r between now and t, we are not friends at r.
This corresponds to the following first-order logic formula.

3t· Time(t) A (t > now) A Apologize(you, t) A

(1.5)
(\:If. Time(f) A (now::::; r < t) => -,Friends(us, t'))
Note that we are explicitly using terms of first-order logic to denote time
points. Using CORA, we can translate (1.4) as follows.

(-,Friends(us)) U Apologize(you) (1.6)

The "u" is a temporal connective of CORA, which is read as "until." The

obvious comment to make is that (1.6) is a much more succinct representation
of (1.4). But more importantly, (1.6) is much closer to the natural language
(1.4) than the rather arcane first-order translation (1.5). Because (1.5) explicitly
incorporates the technical apparatus to deal with time, it has lost the structure
of the English language (1.4), rendering it much harder to read. 3
More generally, modal and temporal logics can be understood as special­
ized languages for representing the properties of relational structures. In the
case of temporal properties, the relational structure in question is the temporal
structure of the world: nodes in the structure correspond to time points, and
arcs between nodes correspond to the passing of time. If we do not use modal
connectives to represent properties of the relational structure, then we seem
obliged to introduce the relational structure into the representation language
itself, as we did with time points in (1.5). As the above example illustrates, this
leads to formulae that are, at the very least, hard to read. In later chapters, we

3 If at this point you do not buy into the argument that (1.6) is more user-friendly than (1.5), then
you are not going to enjoy the rest of this book.
14 Chapter 1

will see how an agent's beliefs, desires, and intentions can also be understood
as relational structures.
With respect to the issue of theorem proving, it should be noted that first­
order logic hardly lends itself to efficient theorem proving. After all, even for
propositional logic the satisfiability problem is Np-complete. Nobody would
argue that theorem proving for modal and temporal logics is computationally
easy. But with time, we will perhaps come to gain a much better understanding
of how to build theorem provers for such logics - just as we have with first­
order logic.

I Am a Logic Guru - Should I Read This Book?

For the purposes of this book, logic is a means to an end and not an end in
itself. In other words, this book is first and foremost about rational agents
and multiagent systems. It is not intended to be a book about logic. So,
although the book contains many formal results, comparatively few of these
are specifically about logic itself. There are no complexity or completeness
results, for example, and no results on the expressive power of the logic used.
Although the book is not primarily about logic, I have nevertheless attempted
to ensure that it is "respectable" from a logical standpoint. 4 Thus the syntax
and semantics of CORA are defined in as much detail as all but the most
fastidious would expect, and detailed proofs are given for all results.
Logics like CORA, which combine multiple modalities into a single
framework, are very much at the cutting edge of contemporary logic re­
search [139]. I therefore hope that this book will provide a fertile source of
ideas and problems for researchers more interested in pure logic than its appli­
cations to agent theory.

I Am a Logic Novice - Should I Read This Book?

If you open a page of this book at random, there is a good chance that you will
be confronted with logical notation of some kind. If you are not familiar with
logic or comfortable with discrete mathematics, then this formalism will no
doubt be off-putting. In fact, the "entry level" for reading this book is actually
quite low. I do not presuppose anything more than some familiarity with basic
discrete math notation (sets, functions, and relations) and the ability to read
formulae of first-order logic. Almost all undergraduate courses in computer

4 A colleague who read a draft of this book quipped that no book about logic in AI could be
respectable these days. If you are inclined to agree, you should probably stop reading now.
Rational Agents 15

science will cover these prerequisites. If you have such a background, then
you should be able to read the book and gain an understanding of the main
ideas without delving into the formal details. Ultimately, all you need to know
to understand most of the book is how to read formulae of CORA. Chapter 3
explains how to read formulae of CORA without going into the details of the
logic at all. If you prefer not to read the mathematically oriented sections of
the book, then you should avoid chapter 4 completely, and all sections marked
with an asterisk ("*" ) in the section header. The appendices survey the logical
preliminaries required to understand the book.

Do I Need to Read the Proofs?

I do not enjoy reading proofs, or indeed writing them, but they are "a necessary
evil" in the words of my editor at MIT Press. If you feel the same as I feel,
then my advice is as follows. If you simply want to become familiar with the
techniques used and the results in the book, then there is no reason to read any
of the proofs at all. However, if you want to use CORA yourself, to build on
any of the results presented in this book, or to gain a really deep understanding
of the book, then you will simply have to bite the bullet. In fact, most of
the proofs rely on a relatively small set of techniques, which will already be
familiar to readers with a grounding in modal, temporal, and first-order logic.

How Do I Implement This Book?

CORA is not an executable logic, and there is currently no simple way of au­
tomating CORA so that it can be used directly as a knowledge representation
formalism or a programming language. The possible roles that logics such as
CORA might play in the engineering of agent systems are explored in detail
in chapter 9.

Where Can I Find Out More?

I have done my best to make the material in this book as self-contained as

possible, in that while it presupposes some basic knowledge of discrete maths
and logic, all definitions and such are included in their entirety. There should
therefore be no need to refer directly to other texts in order to understand the
book. However, to help the reader find out more, every chapter (including this
one) concludes with a section entitled "Notes and Further Reading," which
gives historical notes, pointers to further reading, and open problems. The
appendices also contain background material.
16 Chapter 1

1.6 The Structure of This Book

The remainder of this book can be broadly divided into three parts:

• The first part (chapters 2 and 3) is background material. This part sets the

scene, by introducing the basic concepts and formalism used throughout the
book.
• The second part (chapters 4 and 5) develops the formal model that is used
throughout the book, and shows how this framework can be used to capture
some properties of individual rational agents.
• The third part (chapters 6, 7, and 8) shows how agents can be used to capture

properties of multiagent, social systems.

In more detail:

• Chapter 2 ("The Belief-Desire-Intention Model") gives a detailed introduc­

tion to the 8DI model. It begins by introducing Bratman's intention-based the­

ory of practical reasoning, the foundation upon which the 8DI model rests. Af­
ter introducing this theory, I turn to the question of how 8DI agents might be
implemented. I consider a number of pseudo-code agent control loops, and dis­
cuss the extent to which they can be considered as capturing our pre-theoretic
intuitions about beliefs, desires, and intentions.
• Chapter 3 ("An Introduction to .cORA") informally introduces .cORA.

This chapter is intended for readers who do not have a strong background
in logic. After this chapter, readers should be able to understand formulae of
.cORA.
• Chapter 4 (".cORA Defined") gives a complete definition of the syntax
and semantics of .cORA, and investigates some of its properties. This chapter
presupposes some familiarity with quantified modal, temporal, and dynamic
logics, but is otherwise entirely self-contained.
• Chapter 5 ("Properties of Rational Agents") investigates how .cORA can

be used to capture properties of rational agents. Examples of such properties

include the possible interrelationships among beliefs, desires, and intentions,
and degrees of realism. This chapter focuses in particular on the properties of
individual agents.
• Chapter 6 ("Collective Mental States") changes the focus of the book from

individual agents to groups of agents. It shows how .cORA can be used to for­
malize collective mental states such as mutual beliefs, desires, and intentions,
Rational Agents 17

but in addition, such notions as joint commitment.

• Chapter 7 ("Communication") shows how .cORA can be used to define

communicative (speech) acts between agents. Following a discussion of the

history of speech acts in multiagent systems, I show how "request" and "in­
form" actions can be defined in .cORA, building on the prior work of Cohen
and Levesque [36]. Roughly speaking, an inform action takes place when one
agent attempts to get another agent to believe something; a request action takes
place when one agent gets another agent to intend something. Using these two
speech acts as primitives, I show how a range of other speech acts may be
defined using .cORA.
• Chapter 8 ("Cooperation") shows how .cORA can be used to define a model

of cooperative problem solving. This model is an adaptation of the theory

of cooperation introduced in [247, 248, 250]. It treats cooperative problem
solving as a four-stage process, beginning when one agent recognizes the
potential for cooperation with respect to one of its actions, which is followed
by the agent attempting to solicit assistance from a team of agents, which then
attempt to agree to a plan of Goint) action to achieve the goal, which finally is
executed by the group.
• Chapter 9 ("Logic and Agent Theory") attempts to put the book into per­

spective. It discusses the role that logical theories of agency can or should
play in the development of agent systems. Adopting a software engineering
perspective, where such theories are treated as specifications that an idealized
agent would satisfy, it investigates the extent to which they can be used in the
implementation or verification of practical agent systems.

In addition, the book contains two appendices:

• Appendix A contains a summary of the notation used in the main text of the
book.
• Appendix B contains a short, self-contained introduction to modal and

temporal logics, for readers unfamiliar with this material.

1.7 Notes and Further Reading

There are now many introductions to intelligent agents and multiagent sys­
tems. Ferber [56] is an undergraduate textbook, although as its name suggests,
this volume focussed on multiagent aspects rather than on the theory and prac­
tice of individual agents. A first-rate collection of articles introducing agent
18 Chapter 1

and multiagent systems is WeiB [233]; in particular, chapters 1, 2, 3, and 8

would provide an excellent introduction to the current volume. Three collec­
tions of research articles provide a comprehensive introduction to the field of
autonomous rational agents and multiagent systems: Bond and Gasser's 1988
collection, Readings in Distributed Artificial Intelligence, introduces almost
all the basic problems in the multiagent systems field, and although some of
the papers it contains are now rather dated, it remains essential reading [18];
the Readings in Planning collection, from Allen, Hendler, and Tate may not
appear to be the most obvious place to start for an introduction to agents, but
the basic issue it addresses - deciding what action to perform - is central to
the study of agents [2]; finally, Huhns and Singh's more recent collection sets
itself the ambitious goal of providing a survey of the whole of the agent field,
and succeeds in this respect very well [102]. For a general introduction to the
theory and practice of intelligent agents, see Wooldridge and Jennings [249],
which focuses primarily on the theory of agents, but also contains an extensive
review of agent architectures and programming languages. For a collection of
articles on the applications of agent technology, see [109]. A comprehensive
roadmap of agent technology was published as [108].
Some authors view the whole of the artificial intelligence endeavor as one
of constructing rational agents: Russell and Norvig's enormously successful,
encyclopedic introductory AI textbook is the best-known example [198].
This question of "what is an agent" is one that continues to generate some
debate, particularly on unmoderated email lists and news forums. Unfortu­
nately, much of the debate is not as informed as one might wish. For an in­
teresting collection of answers to the question, see [164]. The notion of a soft­
ware agent, which inhabits a software environment, was introduced by Oren
Etzioni [53]. Another good discussion can be found in Kaelbling [112].
The relevance of decision theory and game or economic theory to artificial
intelligence was recognized in the very earliest days of the discipline; the work
of Herb Simon is perhaps best known in this regard (see, e.g., [208]). Decision
theoretic approaches to the planning problem are currently the focus of con­
siderable attention within the artificial intelligence community - see [16] for
an overview. Rao and Georgeff, who developed the basic BDI logic framework
upon which .cORA is based, investigated the relationship between the BDI
model and classical decision theory in [185, pp.297-200]. They showed how
the "choice and chance" trees employed in decision theory could be mapped
into semantic models for their BDI logics.
Rational Agents 19

Following the seminal work of Jeffrey Rosenschein and colleagues on

the application of game-theoretic techniques to negotiation [195], game the­
ory is now widely applied in multiagent systems research. A comprehensive
overview is provided in [199].
Finally, logic has been widely used in artificial intelligence since John
McCarthy's early work on the "advice taker" system [149]; see [73, 198] for
detailed introductions to the use of logic in artificial intelligence, and see [103]
for a general discussion on the role of logic in artificial intelligence. See
also [167] and Birnbaum's response to this for some more strident views [15].
For discussions on the relative merits of first-order logic versus the alternatives
(modal or temporal logics), see [191] or [70].
Although he is a strong advocate of logic in artificial intelligence, John
McCarthy has reservations about the utility of modal logic; a position paper on
the topic, entitled Modality, Sit Modal Logic, No! appeared as [150]. Joseph
Halpern responded to McCarthy's statement in the online journal Electronic
Transactions on Artificial Intelligence in late 1999; at the time of writing, the
online debate was continuing.
2 The Belief-Desire-Intention Model

In this chapter, I give a detailed introduction to the belief-desire-intention

model of agency. As noted in chapter 1, the 8DI model combines three distinct
components: a philosophical component; a software architecture component;
and a logical component. The purpose of this chapter is primarily to introduce
the first and second components of the 8DI model; the remainder of the book
investigates the logical component in detail. I begin by introducing Bratman's
theory of human practical reasoning.

2.1 Practical Reasoning

The 8DI model has its roots in the philosophical tradition of understanding
practical reasoning in humans. Put simply, practical reasoning is reasoning
directed towards actions - the process of figuring out what to do:

Practical reasoning is a matter of weighing conflicting considerations for and against

competing options, where the relevant considerations are provided by what the agent
desires/values/cares about and what the agent believes. [21, p.17]

It is important to distinguish practical reasoning from theoretical reason­

ing [48]. Theoretical reasoning is directed towards beliefs. To use a rather tired
example, if I believe that all men are mortal, and I believe that Socrates is a
man, then I will usually conclude that Socrates is mortal. The process of con­
cluding that Socrates is mortal is theoretical reasoning, since it affects only
my beliefs about the world. The process of deciding to catch a bus instead of
a train, however, is practical reasoning, since it is reasoning directed towards
action.
Human practical reasoning appears to consist of at least two distinct ac­
tivities. The first of these involves deciding what state of affairs we want to
achieve; the second process involves deciding how we want to achieve these
states of affairs. The former process - deciding what states of affairs to
achieve - is known as deliberation. The latter process - deciding how to
achieve these states of affairs - we call means-ends reasoning.
To better understand deliberation and means-ends reasoning, consider the
following example. When a person graduates from university with a first
degree, he or she is faced with some important choices. Typically, one proceeds
in these choices by first deciding what sort of career to follow. For example,
one might consider a career as an academic, or a career in industry. The process
22 Chapter 2

of deciding which career to aim for is deliberation. Once one has fixed upon a
career, there are further choices to be made; in particular, how to bring about
this career. Suppose that after deliberation, you choose to pursue a career as an
academic. The next step is to decide how to achieve this state of affairs. This
process is means-ends reasoning. The end result of means-ends reasoning is a
plan or recipe of some kind for achieving the chosen state of affairs. For the
career example, a plan might involve first applying to an appropriate university
for a Ph.D. place, and so on. After obtaining a plan, an agent will typically then
attempt to carry out (or execute) the plan, in order to bring about the chosen
state of affairs. If all goes well (the plan is sound, and the agent's environment
cooperates sufficiently), then after the plan has been executed, the chosen state
of affairs will be achieved.
Thus described, practical reasoning seems a straightforward process, and
in an ideal world, it would be. But there are several complications. The first is
that deliberation and means-ends reasoning are computational processes. In all
real agents (and in particular, artificial agents), such computational processes
will take place under resource bounds. By this I mean that an agent will only
have a fixed amount of memory and a fixed processor available to carry out
its computations. Together, these resource bounds impose a limit on the size
of computations that can be carried out in any given amount of time. No real
agent will be able to carry out arbitrarily large computations in a finite amount
of time. Since almost any real environment will also operate in the presence
of time constraints of some kind, this means that means-ends reasoning and
deliberation must be carried out in a fixed, finite number of processor cycles,
with a fixed, finite amount of memory space. From this discussion, we can see
that resource bounds have two important implications:

• Computation is a valuable resource for agents situated in real-time envi­

ronments. The ability to perform well will be determined at least in part by

the ability to make efficient use of available computational resources. In other
words, an agent must control its reasoning effectively if it is to perform well.

• Agents cannot deliberate indefinitely. They must clearly stop deliberating at

some point, having chosen some state of affairs, and commit to achieving this
state of affairs. It may well be that the state of affairs it has fixed upon is not
optimal - further deliberation may have led it to fix upon an another state of
affairs.
The Belief-Desire-Intention Model 23

We refer to the states of affairs that an agent has chosen and committed to
as its intentions. The BDI model of agency is ultimately one that recognizes
the primacy of intentions in practical reasoning, and it is therefore worth
discussing the roles that they play in more detail.

2.2 Intentions in Practical Reasoning

First, notice that it is possible to distinguish several different types of intention.

In ordinary speech, we use the term "intention" to characterize both actions
and states of mind. To adapt an example from Bratman [20, p. 1], I might
intentionally push someone under a train, and push them with the intention
of killing them. Intention is here used to characterize an action - the action
of pushing someone under a train. Alternatively, I might have the intention
this morning of pushing someone under a train this afternoon. Here, intention
is used to characterize my state of mind. In this book, when I talk about
intentions, I mean intentions as states of mind. In particular, I mean future­
directed intentions - intentions that an agent has towards some future state of
affairs.
The most obvious role of intentions is that they are pro-attitudes [2 1, p. 23].
By this, I mean that they tend to lead to action. Suppose I have an intention to
write a book. If I truly have such an intention, then you would expect me to
make a reasonable attempt to achieve it. This would usually involve, at the
very least, me initiating some plan of action that I believed would satisfy the
intention. In this sense, intentions tend to play a primary role in the production
of action. As time passes, and my intention about the future becomes my
intention about the present, then it plays a direct role in the production of
action. Of course, having an intention does not necessarily lead to action. For
example, I can have an intention now to attend a conference later in the year. I
can be utterly sincere in this intention, and yet if! learn of some event that must
take precedence over the conference, I may never even get as far as considering
travel arrangements.
Bratman notes that intentions play a much stronger role in influencing
action than other pro-attitudes, such as mere desires:
My desire to play basketball this afternoon is merely a potential influencer of my
conduct this afternoon. It must vie with my other relevant desires [... ] before it is
settled what I will do. In contrast, once I intend to play basketball this afternoon, the
matter is settled: I normally need not continue to weigh the pros and cons. When the
afternoon arrives, I will normally just proceed to execute my intentions. [21, p.22]
24 Chapter 2

The second main property of intentions is that they persist. If I adopt an

intention to become an academic, then I should persist with this intention and
attempt to achieve it. For if I immediately drop my intentions without devoting
any resources to achieving them, then I will not be acting rationally. Indeed,
you might be inclined to say that I never really had intentions in the first place.
Of course, I should not persist with my intention for too long - if it
becomes clear to me that I will never become an academic, then it is only
rational to drop my intention to do so. Similarly, if the reason for having an
intention goes away, then it would be rational for me to drop the intention. For
example, if! adopted the intention to become an academic because I believed it
would be an easy life, but then discover that this is not the case (e.g., I might be
expected to actually teach!), then the justification for the intention is no longer
present, and I should drop the intention.
If I initially fail to achieve an intention, then you would expect me to try
again - you would not expect me to simply give up. For example, if my first
application for a Ph.D. program is rejected, then you might expect me to apply
to alternative universities.
The third main property of intentions is that once I have adopted an inten­
tion, the very fact of having this intention will constrain my future practical
reasoning. For example, while I hold some particular intention, I will not sub­
sequently entertain options that are inconsistent with that intention. Intending
to write a book, for example, would preclude the option of partying every night:
the two are mutually exclusive. This is in fact a highly desirable property from
the point of view of implementing rational agents, because in providing a "fil­
ter of admissibility," intentions can be seen to constrain the space of possible
intentions that an agent needs to consider.
Finally, intentions are closely related to beliefs about the future. For ex­
ample, if! intend to become an academic, then I should believe that, assuming
some certain background conditions are satisfied, I will indeed become an aca­
demic. For if I truly believe that I will never be an academic, it would be non­
sensical of me to have an intention to become one. Thus if I intend to become
an academic, I should at least believe that there is a good chance I will indeed
become one. However, there is what appears at first sight to be a paradox here.
While I might believe that I will indeed succeed in achieving my intention, if
I am rational, then I must also recognize the possibility that I canfail to bring
it about - that there is some circumstance under which my intention is not
satisfied.
The Belief-Desire-Intention Model 25

From this discussion, we can identify the following closely related situa­
tions:

• Having an intention to bring about <p, while believing that you will not

bring about <p is called intention-belief inconsistency, and is not rational (see,
e.g., [20, pp.37-38]).
• Having an intention to achieve <p without believing that <p will be the case

is intention-belief incompleteness, and is an acceptable property of rational

agents (see, e.g., [20, p.38]).

The distinction between these two cases is known as the asymmetry thesis [20,
pp.37--4 1].
Summarizing, we can see that intentions play the following important roles
in practical reasoning:

• Intentions drive means-ends reasoning.

If! have formed an intention, then I will attempt to achieve the intention, which
involves, among other things, deciding how to achieve it. Moreover, if one
particular course of action fails to achieve an intention, then I will typically
attempt others.
• Intentions persist.
I will not usually give up on my intentions without good reason - they will
persist, typically until I believe I have successfully achieved them, I believe
I cannot achieve them, or I believe the reason for the intention is no longer
present.
• Intentions constrain future deliberation.
I will not entertain options that are inconsistent with my current intentions.
• Intentions influence beliefs upon which future practical reasoning is based.
If I adopt an intention, then I can plan for the future on the assumption that I
will achieve the intention. For if I intend to achieve some state of affairs while
simultaneously believing that I will not achieve it, then I am being irrational.

Notice from this discussion that intentions interact with an agent's beliefs and
other mental states. For example, having an intention to <p implies that I do not
believe <p is impossible, and moreover that I believe given the right circum­
stances, <p will be achieved. However, satisfactorily capturing the interaction
between intention and belief tums out to be surprisingly hard: the way in which
intentions interact with other mental states is considered in chapter 5.
26 Chapter 2

Algorithm: Agent Control Loop Version 1

1. while true
2. observe the world;
3. update internal world model;
4. deliberate about what intention to achieve next;
5. use means-ends reasoning to get a plan for the intention;
6. execute the plan
7. end while

Figure 2.1
A basic agent controi ioop.

2.3 Implementing Rational Agents

Based on the discussion above, let us consider how we might go about building
a BDI agent. The strategy I use is to introduce progressively more complex
agent designs, and for each of these designs, to investigate the type of behavior
that such an agent would exhibit, compared to the desired behavior of a rational
agent as discussed above. This then motivates the introduction of a refined
agent design, and so on.
The first agent design is shown in Figure 2. 1. The agent continually exe­
cutes a cycle of observing the world, deciding what intention to achieve next,
determining a plan of some kind to achieve this intention, and then executing
this plan.
The first point to note about this loop is that we will not be concerned with
stages ( 2) or (3). Observing the world and updating an internal model of it are
important processes, worthy of study in their own right; but they lie outside the
scope of this volume. Instead, we are concerned primarily with stages (4) to (6).
Readers interested in understanding the processes of observing and updating
can find some key references in the "further reading" section at the end of this
chapter.
One of the most important issues raised by this simple agent control loop
follows from the fact that the deliberation and means-ends reasoning processes
are not instantaneous. They have a time cost associated with them. Suppose
that the agent starts deliberating at time to, begins means-ends reasoning at t1,
and subsequently begins executing the newly formed plan at time t2. The time
taken to deliberate is thus

tdeliberate = t1 - to
The Belief-Desire-Intention Model 27

and the time taken for means-ends reasoning is

Further suppose that the deliberation process is optimal in the sense that if it
selects some intention to achieve, then the achievement of this intention at time
to would be the best thing for the agent. So at time t1, the agent has selected an
intention to achieve that would have been optimal if it had been achieved at to.
But unless tdeliberate is vanishingly small, then the agent runs the risk that the
intention selected is no longer optimal by the time the agent has fixed upon it.
Of course, selecting an intention to achieve is only part of the overall
practical reasoning problem; the agent still has to determine how to achieve
the intention, via means-ends reasoning.
As with deliberation, assume that the means-ends reasoning process enjoys
optimality in the sense that if it is given an intention to achieve and it starts
executing at time t1, then it will select the course of action to achieve this
intention such that this course of action would be optimal if executed at time t1.
Again, means-ends reasoning is not instantaneous: it takes time to find the best
course of action to achieve the intention. How long exactly might it take to find
the best action? Suppose the agent has n possible actions available to it at any
given time. Then there are n! possible sequences of these actions. Since n! has
exponential growth, this means that the most obvious technique for finding a
course of action (systematically searching through the space of possible action
sequences) will be impossible in practice for all but the most trivial n. The
problem is exacerbated by the fact that the plan is developed on the basis of an
observation made of the environment at time to. The environment may be very
different by time t2.
Assuming we have an agent with optimal deliberation and means-ends
reasoning components, in the sense described above, it should be clear that
this agent will have overall optimal behavior in the following circumstances:

1. when deliberation and means-ends reasoning take a vanishingly small

amount of time; or
2. when the world is guaranteed to remain static while the agent is deliberating
and performing means-ends reasoning, so that the assumptions upon which the
choice of intention to achieve and plan to achieve the intention remain valid
until the agent has completed deliberation and means-ends reasoning; or
3. when an intention that is optimal when achieved at time to (the time at
which the world is observed) is guaranteed to remain optimal until time t2 (the
28 Chapter 2

time at which the agent has found a course of action to achieve the intention).

Conditions ( 2) and (3) will not hold in most realistic environments. As we

noted in chapter 1, the kinds of environments in which agents work will tend to
be highly dynamic, and such environments will not remain static long enough
for an agent to determine from first principles either an optimal intention or,
given such an intention, an optimal course of action to achieve it. Similarly,
most important tasks that we might assign to agents will be highly time
dependent.
Let us now make the agent algorithm presented above slightly more for­
mal. For this, we require some notation. Throughout this chapter, I make one
important assumption: that the agent maintains some explicit representation of
its beliefs, desires, and intentions. However, I will not be concerned with how
beliefs and the like are represented. One possibility is that they are represented
symbolically, for example as logical statements a III PROLOG facts [34]. In­
deed, for most of this book, it is assumed that beliefs take exactly this form.
However, the assumption that beliefs, desires, and intentions are symbolically
represented is by no means necessary for the remainder of the book. We use
B to denote a variable that holds the agent's current beliefs, and let Bel be the
set of all such beliefs. Similarly, we use D as a variable for desires, and Des
to denote the set of all desires. Finally, the variable I represents the agent's
intentions, and Int is the set of all possible intentions.
Next, we need some way of representing an agent's perception. Consider
a robotic agent, equipped with a video camera, but no other kind of sensor ap­
paratus. Then the video feed delivered to this agent represents the information
available to the agent about its environment. Any representation of its environ­
ment that the robot creates must ultimately be derived from this video feed. We
refer to the video feed as the robot's perceptual input. In implemented agent
systems, perceptual input is usually packaged into discrete bundles, which are
referred to as percepts. We use p, p', PI, . . . to represent percepts that the agent
receives. Let Per be the set of all such percepts.
Finally, we need to say something about plans. Plans and intentions are
closely related. We frequently use the two terms interchangeably in everyday
speech, saying "I plan to. . . " for "I intend to. . . "; as Bratman puts it: "Plans
are intentions writ large" [20, p. 29]. In this book, when we refer to plans, we
are referring to plans as recipes for achieving intentions. For example, one
plan/recipe for achieving my intention of being at the airport might involve
catching a taxi; another might involve catching a bus; and so on. Plans and
The Belief-Desire-Intention Model 29

planning are a major research topic in their own right - see [2] for a detailed
survey. For our purposes, a simple model of plans will suffice. A plan is viewed
as a tuple consisting of:

• a pre-condition, which defines the circumstances under which a plan is

applicable - for example, one of the pre-conditions of the plan to catch a

taxi to the airport is that I have sufficient funds to pay for it;
• a post-condition, which defines what states of affairs the plan achieves -
for example, post-conditions of my plan to catch a taxi to the airport include
me having less money, and me now being located at the airport;
• a body, which is the actual "recipe" part of the plan - for our purposes, the

body is simply a sequence of actions.

We will use 7r (with decorations: 7r', 7rl, ...) to denote plans, and let Plan be
the set of all plans (over some set of actions Ac). We will make use of a number
of auxiliary definitions for manipulating plans (some of these will not actually
be required until later in this chapter):

• if 7r is a plan, then we write pre(7r) to denote the pre-condition of 7r, post( 7r)

to denote the post-condition of 7r, and body(7r) to denote the body of 7r;
• if 7r is plan, then we write empty(7r) to mean that plan 7r is the empty

sequence (thus empty( . . . ) is a boolean-valued function);

• execute( . . . ) is a procedure that takes as input a single plan and executes it

without stopping - executing a plan simply means executing each action in

the plan body in tum;
• if 7r is a plan then by hd(7r) we mean the plan made up of the first action in

the plan body of 7r; for example, if the body of 7r is al, ... , an, then the body
of hd(7r) contains only the action al;
• if 7r is a plan then by tail(7r) we mean the plan made up of all but the first

action in the plan body of 7r; for example, if the body of 7r is al, a2, ... , an,
then the body oftail(7r) contains actions a2, ... , an;
• if 7r is a plan, I � Int is a set of intentions, and B � Bel is a set of beliefs,

then we write sound(7r,1, B) to mean that 7r is a sound plan for achieving I

given beliefs B. (We will not discuss or attempt to define what makes a plan
sound here - the classic paper on the subject is Lifschitz [136].)

We can now define the components of an agent's control loop. An agent's belief
update process is formally modeled as a belief revision function. Such a belief
30 Chapter 2

revision function has the following signature.

brf : p(Bel) x Per -+ p(Bel)

In other words, on the basis of the current beliefs and current percept, the belief
revision function determines a new set of beliefs. (As noted above, in this book
we are not concerned with how belief revision might work; see [71].)
The agent's deliberation process is given by a function

deliberate: p(Bel) -+ p(Int)

which takes a set of beliefs and returns a set of intentions - those selected by
the agent to achieve, on the basis of its beliefs.
An agent's means-ends reasoning is represented by a function

plan: p(Bel) x p(Int) -+ Plan

which on the basis of an agent's current beliefs and current intentions, deter­
mines a plan to achieve the intention. Note that there is nothing in the definition
of the plan(. . . ) function which requires an agent to engage in plan generation
- constructing a plan from scratch [2]. In most BDI systems, the plan( . . . )
function is implemented by giving the agent a plan library [78]. A plan li­
brary is a pre-assembled collection of plans, which an agent designer gives to
an agent. Finding a plan to achieve an intention then simply involves a sin­
gle pass through the plan library to find a plan that, when executed, will have
the intention as a post-condition, and will be sound given the agent's current
beliefs. In implemented BDI agents, pre- and post-conditions are often rep­
resented as (lists of) atoms of first-order logic, and beliefs and intentions as
ground atoms of first-order logic. Finding a plan to achieve an intention then
reduces to finding a plan whose pre-condition unifies with the agent's beliefs,
and whose post-condition unifies with the intention.
The agent control loop is now as shown in Figure 2. 2. This algorithm high­
lights some limitations of this simple approach to agent control. In particu­
lar, steps (4) to (7) inclusive implicitly assume that the environment has not
changed since it was observed at stage (3). Assuming that the time taken to
actually execute the plan dominates the time taken to revise beliefs, deliberate,
or plan, then the crucial concern is that the environment might change while
the plan is being executed. The problem is that the agent remains committed to
the intention it forms at step (5) until it has executed the plan in step (7). If the
environment changes after step (3), then the assumptions upon which this plan
The Belief-Desire-Intention Model 31

Algorithm: Agent Control Loop Version 2

1. B:=Bo; /* Bo are initial beliefs * /
2. while true do
3. get next percept p;
4. B:=bif(B,p) ;
5. I:=deliberate(B) ;
6. 11" :=plan(B,I) ;
7 • execute( 11" )
B. end while

Figure 2.2
A first refinement of the agent control loop.

depends may well be invalid by the time the plan is actually executed.

2.4 The Deliberation Process

So far, we have glossed over the problem of exactly how an agent might go
about deliberating. In this section, we consider the process in more detail. It
is not hard to see that in real life, deliberation typically begins by trying to
understand what the options available to you are. Returning to the career choice
example introduced above, if you gain a good first degree, then one option is
that of becoming an academic; if you fail to obtain a good degree, this option is
not available to you. Another option is entering industry. After deciding what
the options are, you must choose between them, and commit to some. These
chosen options then become intentions.
From this discussion, we can see that the deliberate function as discussed
above can be decomposed into two distinct functional components:

• option generation - in which the agent generates a set of possible alterna­

tives; and
• filtering - in which the agent chooses between competing alternatives, and

commits to achieving them.

We represent option generation via a function, options, which takes the agent's
current beliefs and current intentions, and from them determines a set of
options that we will hereafter refer to as desires. The intuitive interpretation of
a desire is that, in an "ideal world," an agent would like all its desires achieved.
In any moderately realistic scenario, however, an agent will not be able to
32 Chapter 2

Algorithm: Agent Control Loop Version 3

1.
2. B:=Bo; j* Bo are initial beliefs * j
3. 1:=10; /* 10 are initial intentions * /
4. while true do
5. get next percept p;
6. B:=bif(B,p) ;
7. D:=options(B,1) ;
B. 1:=filter(B,D, 1) ;
9. 11" :=plan(B,1) ;
10. execute( 11" )
11. end while

Figure 2.3
Refining the deliberation process into option generation and filtering.

achieve all its desires. This is because desires are often mutually exclusive. For
example, in the OASIS air-traffic control system, which was implemented using
a BDI architecture, an agent was tasked with the problem of finding the optimal
sequence in which to land aircraft at an airport [138]. The option generation
process in OASIS might generate the options of landing two different aircraft
on the same runway at the same time. Clearly, such options are mutually
exclusive: it would be undesirable to land both aircraft simultaneously.
Formally, the signature of the option generation function options is as
follows.

options: p(Bel) x p(Int) --t p(Des)

In order to select between competing options, an agent uses afilter function.

Intuitively, the filter function must simply select the "best" option for the agent
to commit to. We represent the filter process through a function filter, with a
signature as follows.

filter: p(Bel) x p(Des) x p(Int) --t p(Int)

The agent control loop incorporating explicit deliberation and means-ends

reasoning is shown in Figure 2.3. Notice that desires are an input to the filter
process, whereas intentions are an outputfrom it.
The Belief-Desire-Intention Model 33

2.5 Commitment Strategies

When an option successfully passes through the filter function and is hence
chosen by the agent as an intention, we say that the agent has made a commit­
ment to that option. Commitment implies temporal persistence an intention,
-

once adopted, should not immediately evaporate. A critical issue is just how
committed an agent should be to its intentions. That is, how long should an
intention persist? Under what circumstances should an intention vanish?
To motivate the discussion further, consider the following scenario:

Some time in the not-so-distant future, you are having trouble with your ne w household
robot. You say "Willie, bring me a beer." The robot replies "OK boss." Twenty minutes
later, you screech "Willie, why didn't you bring me that beer?" It ans wers "Well, I
intended to get you the beer, but I decided to do something else." Miffed, you send
the wise guy back to the manufacturer, complaining about a lack of commitment. After
retrofitting, Willie is returned, marked "Model C: The Committed Assistant." Again,
you ask Willie to bring you a beer. Again, it accedes, replying "Sure thing." Then you
ask: "What kind of beer did you buy?" It ans wers: "Genessee." You say "Never mind."
One minute later, Willie trundles over with a Genessee in its gripper. This time, you
angrily return Willie for overcommitment. After still more tinkering, the manufacturer
sends Willie back, promising no more problems with its commitments. So, being a
somewhat trusting customer, you accept the rascal back into your household, but as a
test, you ask it to bring you your last beer. Willie again accedes, saying "Yes, Sir." (Its
attitude problem seems to have been fixed.) The robot gets the beer and starts to wards
you. As it approaches, it lifts its arm, wheels around, deliberately smashes the bottle,
and trundles off. Back at the plant, when interrogated by customer service as to why it
had abandoned its commitments, the robot replies that according to its specifications,
it kept its commitments as long as required - commitments must be dropped when
fulfilled or impossible to achieve. By smashing the bottle, the commitment became
unachievable. [35, pp.213-214]

The mechanism an agent uses to determine when and how to drop inten­
tions is known as a commitment strategy. The following three commitment
strategies are commonly discussed in the literature of rational agents [187]:

• Blind commitment
A blindly committed agent will continue to maintain an intention until it
believes the intention has actually been achieved. Blind commitment is also
sometimes referred to asfanatical commitment.
• Single-minded commitment
A single-minded agent will continue to maintain an intention until it believes
that either the intention has been achieved, or else that it is no longer possible
34 Chapter 2

Algorithm: Agent Control Loop Version 4

1.
2. B:=Bo; j* Bo are initial beliefs * j
3. /:=/0; /* 10 are initial intentions * /
4. while true do
5. get next percept p;
6. B:=bif(B,p) ;
7. D:=options(B,1) ;
B. 1:=filter(B,D, 1) ;
9. '/I":=plan(B,1) ;
10. while not empry('/I") do
11. "':=hd('/I") ;
12. execute( "') ;
13. '/1":=tail('/I") ;
14. get next percept p;
15. B:=bif(B,p) ;
16. if not sound('/I",1,B) then
17. '/I":=plan(B,I)
lB. end-if
19. end-while
20. end-while

Figure 2.4
Introducing reactivity: the agent will engage in means ends reasoning when it believes a plao has
(for whatever reason) failed.

to achieve the intention.

• Open-minded commitment
An open-minded agent will maintain an intention as long as it is still believed
possible.

Note that an agent has commitment both to ends (i.e., the state of affairs it
wishes to bring about), and means (i.e., the mechanism via which the agent
wishes to achieve the state of affairs). Currently, our agent control loop is over­
committed, both to means and ends. It never stops to reconsider its intentions,
and it remains committed to plans until they have been fully executed. We will
now see how this basic control loop can be refined in various ways, in order to
obtain different types of commitment.
The first modification we make allows the agent to replan if ever a plan
goes wrong. This reduces the commitment that an agent has towards the means
to achieve its intentions. The revised control loop is illustrated in Figure 2.4.
Using this loop, the agent will be committed to a plan to achieve its
intentions only while it believes that the plan is sound given its beliefs about
The Belief-Desire-Intention Model 35

Algorithm: Agent Control Loop Version 5

1.
2. B:=Bo; j* Bo are initial beliefs *j
3. 1:=10; /* 10 are initial intentions * /
4. while true do
5. get next percept p;
6. B:=bif(B,p) ;
7. D:=options(B,1) ;
B. 1:=filter(B,D, I) ;
9. 11":=plan(B,I) ;
10. while not (empty(1I") or succeeded(I,B) or impossible(I,B) ) do
11. Ci :=hd(1I") ;
12. execute( Ci ) ;
13. 11":=tail(1I") ;
14. get next percept p;
15. B:=bif(B,p) ;
16. if not sound(1I",I,B) then
17. 11":=plan(B,1)
lB. end-if
19. end-while
20. end-while

Figure 2.5
Dropping intentions that are either impossible or that have succeeded.

the current state of the world. If it ever detennines that its plan is no longer
appropriate in order to achieve the current intention, then it engages in further
means-ends reasoning in order to find an alternative plan. Given that its beliefs
are updated every time it executes an action, this implies at least some degree
of reactivity.
This version of the control loop is clearly more attuned to the environment
than previous versions, but it still remains overcommitted to intentions. This
is because although the agent will replan if ever it believes the plan is no
longer appropriate, it never stops to consider whether or not its intentions are
appropriate. This consideration motivates the next version of the control loop,
in which an agent explicitly stops to detennine whether or not its intentions
have succeeded or whether they are impossible. We write succeeded(I, B) to
mean that given beliefs B, the intentions I can be regarded as having been
satisfied. Similarly, we write impossible(I, B) to mean that intentions I are
impossible given beliefs B. The revised control loop is then illustrated in
Figure 2.5.
It is easy to see that this revised agent control loop implements single­
minded commitment. The modifications required to implement open-minded
36 Chapter 2

commitment is straightforward, and is left to the reader. It should be stressed

that, insofar as we are concerned in this book, there is no one "ideal" commit­
ment strategy. Different circumstances call for different commitment strate­
gies.

2.6 Intention Reconsideration

In our current algorithm, an agent will get to reconsider its intentions once
every time around the outer control loop. This implies that it will reconsider its
intentions when one of the following three conditions arises:

• it has completely executed a plan to achieve its current intentions; or

• it believes it has achieved its current intentions; or
• it believes its current intentions are no longer possible.

Although this ensures that the agent is neither undercommitted nor overcom­
mitted to its intentions, it is limited in the way that it permits an agent to re­
consider its intentions. The main problem is that it does not allow an agent to
exploit serendipity. To see what I mean by this, consider the following scenario.
Sophie is a BDI soft ware agent whose task is to obtain documents on behalf of a user.
One day, the user instructs Sophie to obtain a soft copy of the Ph.D. thesis by A. N.
Other, and Sophie creates an intention to this effect. Sophie believes that the Ph.D. is
resident at A. N. Other's www site, and so generates an intention to do wnload it from
there. While she is planning ho w to achieve this intention, ho wever, she is told that a
local copy of the thesis exists. It would clearly be more efficient to obtain this version,
but since there is nothing wrong with the intention of do wnloading the Ph.D. remotely
(she believes it will succeed), she continues to do so.

In this scenario, Sophie would do better to reconsider her intentions while she
is executing the plan. A first attempt to modify the agent control loop would
involve reconsidering intentions every time the agent executed the inner loop
in Figure 2.5: see Figure 2.6. Such an agent is cautious, in the sense that it
always stops to reconsider its intentions before performing an action.
If option generation and filtering were computationally cheap processes,
then this would be an acceptable strategy. Unfortunately, we know that delib­
eration is not cheap - it takes a considerable amount of time. While the agent
is deliberating, the environment in which the agent is working is changing,
possibly rendering its newly formed intentions irrelevant.
The Belief-Desire-Intention Model 37

Algorithm: Agent Control Loop Version 6

1.
2. B:=Bo; j* Bo are initial beliefs *j
3. 1:=10; /* 10 are initial intentions * /
4. while true do
5. get next percept p;
6. B:=brf(B,p) ;
7. D:=options(B,1) ;
B. 1:=filter(B,D, I) ;
9. 11":=plan(B,I) ;
10. while not (empty(1I") or succeeded(I,B) or impossible(I,B) ) do
11. a:=hd(1I") ;
12. execute(a) ;
13. 11":=tail(1I") ;
14. get next percept p;
15. B:=brf(B,p) ;
16. D:=options(B,1) ;
17. 1:=filter(B,D, 1) ;
lB. if not sound(1I",I,B) then
19. 11" :=plan(B,1)
20. end-if
21. end-while
22. end-while

Figure 2.6
A cautious agent, which stops to reconsider intentions before perfonning aoy action.

We are thus presented with a dilemma:

• an agent that does not stop to reconsider its intentions sufficiently often will

continue attempting to achieve its intentions even after it is clear that they
cannot be achieved, or that there is no longer any reason for achieving them;
• an agent that constantly reconsiders its attentions may spend insufficient
time actually working to achieve them, and hence runs the risk of never actually
achieving them.

There is clearly a trade-off to be struck between the degree of commitment and

reconsideration at work here. To try to capture this trade-off, we can modify
the agent to incorporate an explicit meta-level control component. The idea
is to have a boolean-valued function, reconsider, such that reconsider(I, B)
evaluates to "true" just in case it is appropriate for the agent with beliefs B and
intentions I to reconsider its intentions. The agent control loop incorporating
the reconsider( . . . ) function is shown in Figure 2.7.
38 Chapter 2

Algorithm: Agent Control Loop Version 7

1.
2. B:=Bo; j* Bo are initial beliefs * j
3. /:=/0; /* 10 are initial intentions * /
4. while true do
5. get next percept p;
6. B:=bif(B,p) ;
7. D:=options(B,1) ;
8. 1:=filter(B,D, 1) ;
9. 11":= plan(B, 1) ;
10. while not (empty(1I") or succeeded(I,B) or impossible(I,B) ) do
11. CY. :=hd(1I") ;
12. execute( CY. ) ;
13. 11":=tail(1I") ;
14. get next percept p;
15. B:=bif(B,p) ;
16 . if reconsider(I,B) then
17. D:=oplions(B,I) ;
18. 1:=filter(B,D, I) ;
19. end-if
20. if not sound(1I",1,B) then
21. 11":=plan(B,1)
22. end-if
23. end-while
24. end-while

Figure 2.7
An agent that attempts to strike a balance between boldness and caution: whether or not the agent
chooses to reconsider intentions is determined by the booleao-valued function reconsider(...) .

In this version of the agent control loop, the burden of deciding when to
expend effort by deliberating lies with the function reconsider( . . . ) . It is in­
teresting to consider the circumstances under which this function can be said
to behave optimally. Suppose that the agent's deliberation and plan generation
functions are in some sense perfect: that deliberation always chooses the "best"
intentions (however that is defined for the application at hand), and planning al­
ways produces an appropriate plan. Further suppose that time expended always
has a cost - the agent does not benefit by doing nothing. Then it is not diffi­
cult to see that the function reconsider( . . . ) will be behaving optimally if, and
only if, whenever it chooses to deliberate, the agent changes intentions [25 1].
For if the agent chose to deliberate but did not change intentions, then the ef­
fort expended on deliberation was wasted. Similarly, if an agent should have
changed intentions, but failed to do so, then the effort expended on attempting
to achieve its intentions was also wasted.
The Belief-Desire-Intention Model 39

Table 2.1
Practical reasoning situations (cf. [22, p.353]).

Situation Chose to Changed Would have reconsider(...)

number deliberate? intentions? changed intentions? optimal?
1 No No Yes
2 No Yes No
3 Yes No No
4 Yes Yes Yes

The possible interactions between meta-level control and deliberation are

summarized in Table 2. 1:

• In situation ( 1), the agent did not choose to deliberate, and as a consequence,

did not choose to change intentions. Moreover, if it had chosen to deliberate,

it would not have changed intentions. In this situation, the reconsider( . . . )
function is behaving optimally.
• In situation ( 2), the agent did not choose to deliberate, but if it had done so, it
would have changed intentions. In this situation, the reconsider( . . . ) function
is not behaving optimally.
• In situation (3), the agent chose to deliberate, but did not change intentions.

In this situation, the reconsider( . . . ) function is not behaving optimally.

• In situation (4), the agent chose to deliberate, and did change intentions. In

this situation, the reconsider(. . . ) function is behaving optimally.

Notice that there is an important assumption implicit within this discussion:

that the cost of executing the reconsider( . . . ) function is much less than the cost
of the deliberation process itself. Otherwise, the reconsider(. . . ) function could
simply use the deliberation process as an oracle, running it as a subroutine and
choosing to deliberate just in case the deliberation process changed intentions.
The nature of the trade-off was examined by David Kinny and Michael
Georgeff in a number of experiments carried out using a BDl agent sys­
tem [1 16]. The aims ofKinny and Georgeff's investigation were to:

( 1) assess the feasibility of experimentally measuring agent effectiveness in a simulated

environment, ( 2) investigate ho w commitment to goals contributes to effective agent be­
havior and ( 3) compare the properties of different strategies for reacting to change [116,
p.82]

In Kinny and Georgeff's experiments, two different types of reconsideration

strategy were used: bold agents, which never pause to reconsider their inten­
tions before their current plan is fully executed, and cautious agents, which
40 Chapter 2

stop to reconsider after the execution of every action. These characteristics are
defined by a degree of boldness, which specifies the maximum number of plan
steps the agent executes before reconsidering its intentions. Dynamism in the
environment is represented by the rate of world change, 'Y. Put simply, the rate
of world change is the ratio of the speed of the agent's control loop to the rate
of change of the world. If'Y 1, then the world will change no more than once
=

for each time the agent can execute its control loop. If'Y 2, then the world
=

can change twice for each pass through the agent's control loop, and so on. The
performance of an agent is measured by the ratio of number of intentions that
the agent managed to achieve to the number of intentions that the agent had at
any time. Thus if effectiveness is 1, then the agent achieved all its intentions.
If effectiveness is 0, then the agent failed to achieve any of its intentions. The
main results of Kinny and Georgeff's experiments are shown in Figure 2. 8. 1
This graph shows the effectiveness of an agent on the y axis against the dy­
namism of the world (log scale) on the x axis. The key results of Kinny and
Georgeff were as follows.

• If'Y is low (i.e., the environment does not change quickly), then bold agents

do well compared to cautious ones. This is because cautious ones waste time
reconsidering their commitments while bold agents are busy working towards
- and achieving - their intentions.
• If'Y is high (i.e., the environment changes frequently), then cautious agents
tend to outperform bold agents. This is because they are able to recognize when
intentions are doomed, and also to take advantage of serendipitous situations
and new opportunities when they arise.

The bottom line is that different environment types require different intention
reconsideration and commitment strategies. In static environments, agents that
are strongly committed to their intentions will perform well. But in dynamic
environments, the ability to react to changes by modifying intentions becomes
more important, and weakly committed agents will tend to outperform bold
agents.

2.7 Mental States and Computer Programs

We have been talking about programs that have a "belief-desire-intention"

architecture. This is mentalistic terminology, and it is worth pausing for a

1 I am deeply indebted to Martijn Schut for providing this graph.

The Belief-Desire-Intention Model 41

Bold -
Cautious ----.
0.9

0.8

0.7

0.6
..
..
.,
"
.,
.<! 0.5
U

� 0.4

0.3

0.2

0.1

0.6 0.8 1 1.2 1.4 1.6 1.8 2

dynamism (log scale)

Figure 2.8
Kinny and Georgeff's intention reconsideration experiments.

moment to examine the justification for using such terms. It may seem strange
to think of computer programs in terms of mental states such as beliefs, desires,
and intentions. Is it either useful or legitimate to use mentalistic terminology to
characterize machines? As it turns out, there is an extensive literature on this
subject.
When explaining human activity, it is often useful to make statements such
as the following:

Janine took her umbrella because she believed it was going to rain.
Michael worked hard because he wanted to finish his book.

These statements make use of afolk psychology, by which human behavior is

predicted and explained through the attribution of attitudes, such as believing
and wanting (as in the above examples), hoping, fearing, and so on (see
e.g., [2 26, p.l] for a discussion of folk psychology). This folk psychology is
well established: most people reading the above statements would say they
found their meaning entirely clear, and would not give them a second glance.
The attitudes employed in such folk psychological descriptions are called
42 Chapter 2

the intentional notions. 2 The philosopher Daniel Dennett has coined the term
intentional system to describe entities "whose behavior can be predicted by
the method of attributing belief, desires and rational acumen" [43, p.49], [42].
Dennett identifies different "levels" of intentional system:

Afirst-order intentional system has beliefs and desires (etc.) but no beliefs and desires
about beliefs and desires. [... ] A second-order intentional system is more sophisti­
cated; it has beliefs and desires (and no doubt other intentional states) about beliefs and
desires (and other intentional states) -both those of others and its o wn. [43, p.243]

One can carry on this hierarchy of intentionality as far as required.

Now we have been using phrases like belief, desire, intention to talk
about computer programs. An obvious question is whether it is legitimate
or useful to attribute beliefs, desires, and so on to artificial agents. Isn't this
just anthropomorphism? McCarthy, among others, has argued that there are
occasions when the intentional stance is appropriate:

To ascribe belieJs, Jree will, intentions, consciousness, abilities, or wants to a machine

is legitimate when such an ascription expresses the same information about the machine
that it expresses about a person. It is useful when the ascription helps us understand the
structure of the machine, its past or future behavior, or ho w to repair or improve it. It
is perhaps never logically required even for humans, but expressing reasonably briefly
what is actually kno wn about the state of the machine in a particular situation may
require mental qualities or qualities isomorphic to them. Theories of belief, kno wledge
and wanting can be constructed for machines in a simpler setting than for humans,
and later applied to humans. Ascription of mental qualities is most straightfor ward for
machines of kno wn structure such as thermostats and computer operating systems, but
is most useful when applied to entities whose structure is incompletely known. [148],
(quoted in Shoham [205]; underlining is from [205])

What objects can be described by the intentional stance? As it turns out, almost
any automaton can. For example, consider a light switch:

It is perfectly coherent to treat a light s witch as a (very cooperative) agent with the
capability of transmitting current at will, who invariably transmits current when it
believes that we want it transmitted and not otherwise; flicking the s witch is simply
our way of communicating our desires. [205, p.6]
And yet most adults in the modem world would find such a description absurd
- perhaps even infantile. Why is this? The answer seems to be that while
the intentional stance description is perfectly consistent with the observed

2 Unfortunately, the word "intention" is used in several different ways in logic and the philosophy
of mind. First, there is the BDI-like usage, as in "I intended to kill him." Second, an intentional
notion is one of the attitudes, as above. Finally, in logic, the word intension (with an "s") means the
internal content of a concept, as opposed to its extension. In what follows, the intended meaning
should always be clear from context.
The Belief-Desire-Intention Model 43

behavior of a light switch, and is internally consistent,

...it does not buy us anything, since we essentially understand the mechanism suffi­
ciently to have a simpler, mechanistic description of its behavior. [205, p.6]

Put crudely, the more we know about a system, the less we need to rely
on animistic, intentional explanations of its behavior. 3 An obvious question
is then, if we have alternative, perhaps less contentious ways of explaining
systems, why should we bother with the intentional stance? Consider the
alternatives available to us. One possibility is to characterize the behavior of
a complex system by using the physical stance [44, p.36]. The idea of the
physical stance is to start with the original configuration of a system, and then
use the laws of physics to predict how this system will behave:

When I predict that a stone released from my hand will fall to the ground, I am using
the physical stance. I don't attribute beliefs and desires to the stone; I attribute mass, or
weight, to the stone, and rely on the la w of gravity to yield my prediction. [44, p.37]

Another alternative is the design stance. With the design stance, we use knowl­
edge of what purpose a system is supposed to fulfill in order to predict how it
behaves. Dennett gives the example of an alarm clock [44, pp.37-39]. When
someone presents us with an alarm clock, we do not need to make use of phys­
ical laws in order to understand its behavior. We can simply make use of the
fact that all alarm clocks are designed to wake people up if we set them with
a time. No understanding of the clock's mechanism is required to justify such
an understanding - we know that all alarm clocks have this behavior.
However, with very complex systems, even if a complete, accurate pic­
ture of the system's architecture and working is available, a physical or design
stance explanation of its behavior may not be practicable. Consider a computer.
Although we might have a complete technical description of a computer avail­
able, it is hardly practicable to appeal to such a description when explaining
why a menu appears when we click a mouse on an icon. In such situations,
it may be more appropriate to adopt an intentional stance description, if that
description is consistent, and simpler than the alternatives.

3 Shoham observes that the move from an intentional stance to a technical description of behavior
correlates well with Piaget's model of child development, and with the scientific development
of humankind generally [205]. Children will use animistic explanations of objects - such as
light switches - until they grasp the more abstract technical concepts involved. Similarly, the
evolution of science has been marked by a gradual move from theological/animistic explanations
to mathematical ones. The author's own experiences of teaching computer programming suggest
that, when faced with completely unknown phenomena, it is not only children who adopt animistic
explanations. Indeed, it often seems easier to teach some computer concepts by using explanations
such as: "the computer doesn't know. . . ," than to try to teach abstract principles first.
44 Chapter 2

Note that the intentional stance is, in computer science terms, nothing
more than an abstraction tool. It is a convenient shorthand for talking about
complex systems, which allows us to succinctly predict and explain their
behavior without having to understand how they actually work. Now, much of
computer science is concerned with looking for good abstraction mechanisms,
since these allow system developers to manage complexity with greater ease.
The history of programming languages illustrates a steady move away from
low-level machine-oriented views of programming towards abstractions that
are closer to human experience. Procedural abstraction, abstract data types,
and most recently, objects are examples of this progression. So, why not
use the intentional stance as an abstraction tool in computing - to explain,
understand, and, crucially, program complex computer systems?
There are other reasons for believing that an intentional stance will be use­
ful for understanding and reasoning about computer programs [10 2]. First, and
perhaps most importantly, the ability of heterogeneous, self-interested agents
to communicate seems to imply the ability to talk about the beliefs, aspirations,
and intentions of individual agents. For example, in order to coordinate their
activities, agents must have information about the intentions of others [105].
This idea is closely related to Newell's knowledge level [166]. Later in this
book, we will see how mental states such as beliefs, desires, and the like are
used to give a semantics to speech acts [20 2, 35]. Second, mentalistic models
are a good candidate for representing information about end users. For exam­
ple, imagine a tutoring system that works with students to teach them JAVA
programming. One way to build such a system is to give it a model of the
user. Beliefs, desires, and intentions seem appropriate for the makeup of such
models.
For many researchers in AI, this idea of programming computer systems
in terms of "mentalistic" notions such as belief, desire, and intention is the
key component of agent-based computing. The idea of programming computer
systems in terms of mental states was articulated most clearly by Yoav Shoham
in his agent-oriented programming (AOP) proposal [206]. BDI systems can be
viewed as a kind of AOP.

2.8 Notes and Further Reading

Some reflections on the origins of the BDI model, and on its relationship to
other models of agency, may be found in [74]. Belief-desire-intention archi-
The Belief-Desire-Intention Model 45

tectures originated in the work of the Rational Agency project at Stanford Re­
search Institute in the mid- 1980s. Key figures were Michael Bratman, Phil
Cohen, Michael Georgeff, David Israel, Kurt Konolige, and Martha Pollack.
The origins of the model lie in the theory of human practical reasoning de­
veloped by the philosopher Michael Bratman [20], which focuses particularly
on the role of intentions in practical reasoning. The conceptual framework of
the 8DI model is described in [2 2], which also describes a specific 8DI agent
architecture called IRM A .
The best-known implementation of the 8 D I model is the PRS system, devel­
oped by Georgeff and colleagues [78, 77]. The PRS has been re-implemented
several times since the mid- 1980s, for example in the Australian AI Institute's
DMARS system [46], the University of Michigan's c + + implementation U M ­
PRS, and a JAVA version called JAM ! [100]. JACK is a commercially available
programming language, which extends the JAVA language with a number of
8DI features [26].
The description of the 8DI model given here draws upon [22] and [188],
but is not strictly faithful to either. The most obvious difference is that I
do not incorporate the notion of the "filter override" mechanism described
in [2 2], and I also assume that plans are linear sequences of actions (which
is a fairly "traditional" view of plans), rather than the hierarchically structured
collections of goals used by PRS .
Plans are central to the 8DI model of agency. An excellent discussion on
the 8DI model, focusing in particular on the role of plans in practical reasoning,
is Martha Pollack's 199 1 Computers and Thought award lecture, presented at
the IJCAI-9 1 conference in Sydney, Australia, and published as "The Uses of
Plans" [180]. Another article, which focuses on the distinction between "plans
as recipes" and "plans as mental states" is [179].
It is worth emphasizing that the 8DI model is only one solution to the prob­
lem of building autonomous rational agents. Many other software architectures
for agent systems have been described in the literature [249, 24]. The so-called
"layered" architectures are currently very popular [24 2]; examples include Fer­
guson's TOURING MACHINES [57, 58], Muller's INTERRAP [163, 16 2], and the
3T architecture [17]. I mentioned in passing that belief revision is not a concern
of this book: see Gardenfors [71] for further information.
Finally, although in this book I focus on what might be called the "ortho­
dox" 8DI model, the 8DI model forms a central component of many other sys­
tems, which either draw inspiration from it or implement parts of it. Examples
include [6 1, 1 1 1, 158, 26].
3 Introduction to CORA

In this chapter, I present an overview of the logical framework that will be

used throughout the remainder of the book. This logic allows us to represent
the properties of rational agents and reason about them in an unambiguous,
well-defined way. The logic is called .cORA, which stands for ".cogic Of
Rational Agents."
Like any logic, .cORA has a syntax, a semantics, and a proof theory. The
syntax of .cORA defines a set of acceptable constructions known as well­
formed formulae (or just formulae). The semantics assign a precise meaning
to every formula of .cORA. Finally, the proof theory of .cORA tells us some
basic properties of the logic, and how we can establish properties of the logic.
The language of .cORA combines four distinct components:

1. Afirst-order component, which is in essence classical first-order logic [223,

52, 73]. This component allows us to represent the properties of objects within
a system under study, and how these objects stand in relation to one another.
2. A belief-desire-intention component, which allows us to express the beliefs,
desires, and intentions of agents within a system.
3. A temporal component, which allows us to represent the dynamic aspects
of systems - how they vary over time.
4. A n action component, which allows us to represent the actions that agents
perform, and the effects of these actions.

From a logical point of view, combining these elements into a single frame­
work makes .cORA rather complex. For this reason, the presentation of
.cORA is divided into two parts. The first part (this chapter), simply gives
an informal introduction to the logic: the main constructs it provides, the in­
tuitive semantics of these constructs, how these constructs can be used to ex­
press properties of agents and their environments, but most importantly, how to
readandunderstand formulae of .cORA. This chapter is intended for readers
whose logical background is not strong. If you are already familiar with the
principles of first-order, modal, temporal, and dynamic logic, then you may
wish to skim through this chapter and move quickly to the next chapter, which
contains a detailed formal presentation of the language.
48 Chapter 3

Table 3.1
First-order logic connectives and quantifiers.

Formula Interpretation
-' 'P 'P is not true
'P/\1/J 'P is true and 1/J is true
'PV1/J 'P is true or 1/J is true
'P�1/J if 'P is true then 1/J is true
'P{:}1/J 'P is true if, and only if, 1/J is true
( 7 = 7' ) 7 is the same as 7
'

\/x· 'P(x) all objects x have property 'P

:Jx . 'P(x) there exists an object x with property 'P

3.1 The First-Order Component of .cORA

The first thing to say about .cORA is that it is an extension of classical

first-order logic. Thus it contains the usual connectives of first-order logic,
including "--," (not), "/\" (and), "V" (or), "=>" (implies), and "¢?" (if, and only
if) . .cORA also contains the first-order quantifiers "\I" (for all) and ":3" (there
exists) - see Table 3.1.
To recap, let us consider some fonnulae of classical first-order logic, which
are thus also fonnulae of .cORA. We may as well start with an overused
example.

\Ix· Man(x) => Mortal(x) (3.1)

This fonnula simply asserts that all objects x which satisfy property Man also
satisfy the property Mortal. The intended interpretation of the fonnula is that
all men are mortal.
Fonnally, the x in this fonnula is a variable. By default, we will use the
lowercase italic roman letters x, y, and z as variables throughout the book, in
order to give the reader some help in understanding fonnulae.
Man and Mortal are predicates. The Man and Mortal predicates each take
a single argument, and hence they have an arity of one. Predicates with arity
one (i.e., that take a single argument) are called properties, or unary predicates.
By default, predicate symbols are written as strings of roman letters, starting
with an uppercase letter. Where appropriate, predicates are InterCapped in the
interests of readability.
The symbol "\I" is the universal quantifier, and is read as "for all." The
symbol "=>" is the logical connective for implication. Any fonnula with the
shape cp => 'l/J is read as " cp implies 'l/J," or "if cp then 'l/J." Finally, the raised dot
". " and parentheses ") " and "( " are punctuation symbols.
Introduction to .cORA 49

The second example illustrates the existential quantifier, ":3."

\:Ix· Man(x) ::::}:3y . Woman(y) A MotherOf (x,y) (3.2)

This formula asserts that for every object x that satisfies the property Man,
there exists at least one object y such that y satisfies the property of being a
woman, and in addition, the MotherOf relationship exists between x and y. In
other words, for every man x, there is a woman y such that y is the mother of x.
In this example, x andy are variables, Man and Woman are predicate symbols
of arity one, and MotherOf is a predicate of arity two. Predicates of arity two
are often referred to as two-place predicates, or binary predicates.
The third example illustrates the use of constants.

:3x . Monitor(x) A MonitorState(x, rea dy) (3.3)

This formula asserts that there is at least one objectx such that x is a Monitor
and the state of x is rea dy. Here, Monitor and MonitorState are predicates
of arity one and two respectively, x is a variable, and rea dy is a constant.
A constant is just a fixed name for a particular thing. Thus formula (3.3)
intuitively expresses the fact that some monitor is in a "ready" state.

\:Ix· \:Iy . Person(x) A Person(y) ::::} -,Superior(x,y) (3.4)

This formula asserts that for all peoplex and y, x is not superior to y in other
-

words, nobody is superior to someone else. Where we have multiple variables

quantified in the same way appearing in succession, as in this example where
we have the two universally quantified variables x and y, we simplify the
formula as follows.

\:Ix,y . Person(x) A Person(y) ::::} -,Superior(x,y) (3.5)

We also have equality in our language, allowing us to compare terms.

\:Ix,y· USPresident(x) A USPresident(y) ::::} (x = y) (3.6)

This formula asserts that if x is u.s. President and y is u.s. President, then x
andy must be the same. In other words, there is only one u.s. President.
Examples (3.1}-{3.6) demonstrate some conventions that will be used
throughout the remainder of the book:

• predicates will always be written as InterCapped strings of italic roman

letters starting with an uppercase letter;
• variables will be written as the lower-case italic roman letters x, y, and z;
50 Chapter 3

Table 3.2
Belief, desire, and intention modalities.

Formula Interpretation
(Bel i cp) agent i believes cp
(Des i cp) agent i desires cp
(Int i cp) agent i intends cp

• constants will be written as strings of lowercase italic roman letters; and

• parentheses (ordinary brackets, such as those surrounding this text) and
other types of brackets are used where appropriate to make the intended
reading of formulae clear.

Before proceeding, see if you can determine for yourself the intended
interpretation of the following formulae.

• \:Ix· (Vegetarian(x) 1\ -,Drinks(x, beer)) => Skinny(x)

• \:Ix· -, Rich(x) => -,Happy(x)

• \:Ix· Pleasant(x) => (Jllegal(x) V Immoral(x))

3.2 The Belief, Desire, Intention Component of CORA

So far, all we have is ordinary first-order logic. But CORA extends first-order
logic in a number of important ways. First, CORA incorporates a whole class
of extra connectives, known as modal connectives or modalities, which allow
us to represent the beliefs, desires, and intentions of agents - see Table 3.2
for an overview. We will illustrate these modal connectives by way of example
at first, and then discuss some technicalities associated with their use.
We will start with belief. The basic idea is to have a class of formulae with
the following structure.

(Bel i cp) (3.7)

Here, i is a term (i.e., a variable or a constant) that denotes an agent, and cp is

any formula of CORA. Notice that i must be a term standing for an agent. To
give the reader some visual clues when reading formulae, I use i andj as terms
denoting agents. The formula (3.7) is intended to express the fact that the agent
i believes cpo To illustrate the idea, here is a simple formula about beliefs.

(Beljanine Sunny(melbourne)) (3.8)

Introduction to .cORA 51

The intended interpretation of this formula is that the agentjanine believes

that Sunny(melbourne). In other words, Janine believes it is sunny in Mel­
bourne. Here, janine is a first-order logic constant, which denotes a particular
agent; Sunny(melbourne) is simply a formula of CORA.

3 i· (Bel i Sunny(melbourne)) (3.9)

This formula says that some agent i believes it is sunny in Melbourne - in

other words, that somebo dy believes it is sunny in Melbourne. Similarly, we
can say that everyone believes it is sunny in Melbourne.

Vi· (Bel i Sunny(melbourne)) (3.10)

We can constrain the variable i so that it only refers to English people - the
intended reading of (3.11) is thus that every English person believes it is sunny
in Melbourne.

Vi· En glishPerson(i) => (Bel i Sunny(melbourne)) (3.11)

Note that, although Bel is written a little like an ordinary first-order logic
predicate that takes two arguments, it is actually very different. The most
obvious difference is that the arguments to an ordinary first-order predicate
must all be terms of first-order logic: variables, constants, or functions that
return some value. The first argument to the Bel modality is indeed a term,
although it is a special kind of term - one that stands for an agent. To see why
this is important, consider that although the first argument to the Bel modality
in the following formula is a term, the formula does not make much sense.

(Bel 0.5 Sunny(melbourne))

In order to prevent such nonsensical formulae, it is therefore required that the

first argument to the Bel modality is a term that stands for an agent.
The second argument to the Bel modality, however, is certainly not a
term. It is actually a formula of CORA. To emphasize this point, consider
the following formula of CORA - the intended interpretation is that Janine
believes all Australian cities are sunny.

(Beljanine Vx· AustralianCity(x) => Sunny(x)) (3.12)

Just as we must be careful to ensure that the first argument to the Bel modality
is a term denoting an agent, so we must be careful to ensure that the second
argument is a formula. As the second argument in the following formula is a
52 Chapter 3

natural number, (3.13) is not a legal formula of CORA.

(Beljanine 2) (3.13)

We must therefore be careful how we construct belief, desire, and intention

formulae of CORA, to be sure that we are not breaking any logical rules. For
example, although we can quantify over the first argument to the Bel predicate
(the term denoting the agent), we are not allowed to quantify over the second
(the formula). Thus the following is not a legal formula of CORA.

3x . (Beljanine x) (3.14)

You might absentmindedly write this formula while trying to represent the fact
that there is something that Janine believes. As it turns out, statements like this
cannot be represented in our language. The following misguided attempt to
represent the statement "Janine believes everything Michael believes" is also
illegal in CORA.

Vx· (Bel michael x) => (Beljanine x) (3.15)

Although some logics provide the machinery to represent such statements, they
are somewhat complex, and will not be necessary for us. (A short discussion
on the merits and drawbacks of such metalogics is provided in appendix B.)
Beliefs can be nested. The intended interpretation of the following formula
is that Janine believes Michael believes that London is rainy.

(Beljanine (Bel michael Rainy(london))) (3.16)

Try to understand the intended interpretation of the following formula for

yourself.

(Beljanine Vi· Australian(i) => (Bel i Rainy(london))) (3.17)

Of course, belief is just one of three modalities that we use to characterize the
state of 8DI agents - the other two are, not surprisingly, desires and intentions.
We introduce another two modal connectives, (Des i cp) and (Int i cp), to
represent desires and intentions respectively. The rules for making desire and
intention formulae are the same as for making belief formulae. Thus the
following formula is intended to say that Janine desires it to be sunny in
Melbourne.

(Desjanine Sunny(melbourne)) (3.18)

Introduction to .cORA 53

Similarly, the following formula says that Michael intends to complete his
lecture slides.

(Int michael Complete(slides)) (3.19)

As with beliefs, desires and intentions can be nested. Perhaps more interesting,
however, is the fact that we can nest belief, desire, and intention modalities
with one another. For example, consider the following formula, which is in­
tended to express the fact that Michael intends that Janine believes it is sunny
in Melbourne.

(Int michael (Bel)anine Sunny(melbourne))) (3.20)

The following formula says that Michael desires that everyone will intend to
own a book by Michael.

(Desmichael Vi· (Int i OwnBookBy(michael)))) (3.21)

The possible relationships between an agent's beliefs, desires, and intentions,

and in particular, the relationships that a rational agent should enjoy, will be
the focus of much of the remainder of this book.
Before proceeding, see if you can determine for yourself the intended
interpretation of the following formulae:

• Vi,)· MotherOf (i,)) => (Bel) Wonderful(i))

• (Des)anine Vi· (Bel i Dan gerous(ozoneHole)))
• :3i· (Bel i V)· (Bel) Fool(i)))
• :3i· (Int i V)· (Bel) (Int i ...,PrimeMinister(i))))

Before presenting the temporal component, we will deal with some slightly
more advanced issues connected with modalities and quantification. If you
found the material presented so far hard going, you may wish to skip this
section on a first reading.

Substituting Terms in Modal Contexts*

One of the most difficult issues surrounding the combination of modal opera­
tors and first-order logic is connected with the problem ofsubstitutin g terms in
the context of modal connectives. Essentially, the standardsubstitution rules of
first-order logic do not apply in the cont ext of modal operators. To understand
54 Chapter 3

this problem, consider the following example. !

Janine believes Keanu Reeves (an actor) is cute. In addition to being an actor, Keanu
Reeves is also bass-player for the grunge metal band "Dogstar." Although Janine has
heard of Dogstar, and knows they have a bass player, she has no idea that this bass
player is Keanu Reeves. Does Janine believe the bass-player of Dogstar is cute?

We can summarize this state of affairs in the following formula of .cORA.

(Beljanine Cute(keanu)) 1\ (dogstarBassPlayer = keanu) (3.22)

The question is thus: Should it be possible to derive (3.23) from (3.22)?

(Beljanine Cute(dogstarBassPlayer)) (3.23)

Intuition suggests such a derivation is unacceptable - belief simply does not

work that way. We can ask Janine whether the bass-player of Dogstar is cute,
and she will happily tell us that she has no opinion either way. In ordinary
first-order logic, however, the fact that we have do gstarBassPlayer = keanu
would allow us to substitute do gstarBassPlayer for keanu in a formula without
affecting the truth value of the formula. We would not get far manipulating
ordinary algebraic equations if we were not able to substitute equal terms!
However, with belief statements - and also with other modal operators such
as Des and Int - we are not allowed to do this.
To remind ourselves that Bel is not a predicate, but a modal connective, we
use the LISP-like notation (Bel . . . ) rather than the predicate-like Bel( . . . ) , and
in addition we write modal operator names in an upright sans-serif font rather
than the italic roman font used for predicates.

Quantifying in to Modal Contexts*

Combining quantification with modal operators can lead to some interesting (if
somewhat arcane) situations. Consider the following formula, which its author
intended to express the fact that Janine believes somewhere is sunny.

(Beljanine:3x· Sunny(x)) (3.24)

Now contrast (3.24) with the following formula.

:3x . (Beljanine Sunny(x)) (3.25)

1 This is a version of the well-known "morning star" puzzle [54, p.84].

Introduction to .cORA 55

Are these two formulae expressing the same or different properties? In our
framework, they are different. Formula (3.24) expresses aweaker property than
(3.25). Specifically, it is expressing the fact that Janine believes somewhere is
sunny, but that she does not necessarily know where - she is not necessarily
aware of the identity of the sunny place. In contrast, (3.25) expresses a stronger
property. It says that there is a place x such that Janine believes of x that
it is sunny. Implicitly, (3.25) says that Janine is aware of the identity of the
sunny place. The terminology is that in (3.25), the variable x is quantified de
re, whereas in (3.24), it is quantified de dicto. The practical upshot of this is
that we need to be quite careful when quantifYing in to the scope of a modal
connective, to be sure that we get the interpretation we intend.
Quantifying in to modal contexts has long been recognized as one of the
trickier topics in modal logic; see [101, pp.170-21O] and [54, pp.83-86] for
discussions.

3.3 The Temporal Component of CORA

First-order formulae and the modal connectives introduced above allow us to

express the properties of instantaneous states of affairs. For example, when we
write

(Beljanine Sunny(melbourne))

it is implicitly assumed that this formula is expressing a property of some

particular time. At some later time, it may well be that Janine no longer
believes it is sunny in Melbourne. I now introduce a number of connectives
that allow us to express the dynamics of agents and their environments -
how they change over time. In order to do this, I need to say a little about
the temporal model that underpins CORA. We start from the assumption that
the environment may be in any of a set of possible states. One of these states
represents "now." The past, as viewed from "now" is a linear, discrete sequence
of states. In other words, there is only one past. The course of future events,
however, is yet to be determined. From any given time point, there will be a
number of possible sequences of time points that represent possible futures. In
other words, time branches into the future.
Branching time may initially seem at odds with our intuitions about the
nature of the world - common sense suggests that there is really only one
history in the real world. The branching structure is quite natural when we
56 Chapter 3

t5

t8

Figure 3.1
An example of a branching temporal structure.

think of representing the choices available to agents with respect to the actions
they perform. Depending on which choices these agents make, the future may
unfold in different ways. But also, we assume there is inherently some non­
determinism in the environment. That is, the result of an action performed by
an agent is not pre-determined - there will be a number of possible outcomes.
To summarize, while the past is seen as linear, the future is seen as branch­
ing. Moreover, we assume that there is no "end of time." In our model, time
branches infinitely into the future. All this may sound rather abstract unless
you are familiar with temporal logics, that is, logics for reasoning about dy­
namic environments. To make it more concrete, a simple temporal model is
presented in Figure 3.1 (more accurately, Figure 3.1 shows part of a temporal
model).
In this model, there are ten time points, labeled to to t9 respectively. Each
time point is shown as an oval. The arrows between ovals show the progression
of time, and so ovals to the right are "later" than those to the left. We take
"now" - our index into this model - to be t1. Thus the flow of time to the
left of t 1 (the past) is linear, and to the right of t 1 (the future) is branching.
Time points are labeled with the statements that are true in them. Thus
in t1, the statements q, r, and s are all true, and all other predicates are false.
Introduction to .cORA 57

Table 3.3
Path connectives.

Fonnula Interpretation
'P is true next
'P is eventually true
'P is alway s true
'P is true until 'I/J is true
'P is true unless 'I/J is true

(Assume p is shorthand for Sunny(melbourne), and so on.)

From any time point in Figure 3.1, a number of futures are possible. From
t1, for example, the following are all possible futures:
(t1,t2,t5, . . .)
(t1,t2,t6, . . .)
(t1,t3,t7, . . .)
(t1,t4,t8, . . .)
(t1,t4,t9, . . .)
We refer to these futures as paths. The first part of the temporal fragment of
CORA consists of a number of modal connectives that allow us to express
the properties of these paths - see Table 3.3. These path connectives have the
following interpretation:

• "0" means "next" - thus 0cp is satisfied now if cp is satisfied at the next
moment;
• "0" means "sometime" - thus Ocp is satisfied now if cp is satisfied either
now or at some future moment;
• "0" means "always" - thus 0 cp is satisfied now if cp is satisfied now and

at all future moments;

• "U" means "until" - thus cpU'ljJ is satisfied now if'ljJ is satisfied at some

future moment, and cp is satisfied until then - W is a binary connective

similar to U , allowing for the possibility that the second argument might never
be satisfied.

With respect to the path (t1, t2, t6, . . .) it is easy to see that the formula
o(q 1\ r 1\ -,s 1\ -,t) is true, since q and r are both true in time point t2 (i.e., the
next time point along) and s and t are both false in t2. Similarly, the formula
O -,t is true on this path, since t is eventually false on this path (in fact, it is
false in time points t2 and t6). The formula Or is true on this path, because r
is true on every time point of the path. However, the formula Oq is not true
58 Chapter 3

Table 3.4
Path quantifiers.

Formula Interpretation
cp is true on all paths
cp is true on some path

when evaluated on this path, since while q is true in time points t1 and t2, it is
not true in time point t6. Similarly, the formula <It is not true when evaluated
on this path, since t is not true at any time point on this path.
With respect to the path (t1, t4, t9, . . .) , the formula 0(q 1\ r 1\ s) is true,
since q, r, and s are all true in time point t4. The formula qUt is true on this
path, because q is true at all time points until the time point where t is true
(time point t9).
We are allowed to combine temporal operators, so that <I(PUq) is an
allowable formula, for example. We refer to the formulae that may be built
up using these path connectives as path formulae.
The next connectives provided by .cORA allow us to quantifY over paths
- see Table 3.4. The idea is as follows. Consider the five paths that originate
in time point t1. Then the formula

Dr (3.26)

is true on all these paths. In other words, (3.26) is inevitable. We have a modal
connective "A" that allows us to express the fact that some path formula is
inevitable:

ADr

We interpret a formula Arp as "on all paths that emerge from now, rp is true."
We call "A" the universal path quantifier.
You might guess that, as well as having a universal path quantifier, we
have an existential path quantifier, and indeed we do: " E." A formula E rp is
interpreted as "on some path that emerges from now, rp is true." This formula
will be true in some time point t if there is at least one path (possibly more than
one) rooted at t such that the path formula rp is true on this path. Thus in time
point t1, the formula

E(qUt)

is true, since qUt is true on the path (t1,t4,t9, . . .) , which is rooted at t1.
Formulae that have no temporal operators, or that are prefixed with "A" or " E"
Introduction to .cORA 59

operators are known as state formulae.

Before proceeding, you might want to determine for yourself the intended
meaning of the following formulae. See if they are true in the temporal struc­
ture in Figure 3. 1; if not, try to create a structure in which they are true.

• ( AOp) A( E <)-p)
• A[(Op) A (0O -p )]

• A[(Op) A (O -p )]

• A[(P =? q) V (p =? O·r)]

Quantifying in to Temporal Contexts*

We saw above that quantifying in to belief, desire, and intention contexts can
lead to complex situations. The temporal operators of .cORA are also modal
connectives, of a kind, and they also fall prey to these problems. To see what
I mean by this, consider the following (path) formula, which is intended to be
read as "Tony is and always will be Prime Minister."

OPM(tony) (3. 27)

Now consider the following two existentially quantified versions of this for­
mula.

O::lx· PM(x) (3. 28)

::Ix· OPM(x) (3. 29)

Formula (3. 28) asserts that there is always some Prime Minister. That is,
whatever time point we pick, there will be an individual that has the property
of being Prime Minister.
In contrast, (3. 29) asserts that there is some individual who is always Prime
Minister. The distinction is that in (3. 28), there could be a different Prime
Minister at every time point. In contrast, in (3. 29), there is just one individual
who is Prime Minister in every time point. It should be clear that (3. 29) implies
(3. 28), but the converse does not hold.
A similar situation arises when we consider the " <)" operator. Consider the
following example.

<)'t:/x· Free(x) (3. 30)

Formula (3. 30) asserts that at some point in the future, everybody will be
free. In other words, there will be some particular time in the future at which
60 Chapter 3

everybody will simultaneously be free. Now consider the following variation

on this theme.

\:Ix· ¢Free(x) (3.31)

This example asserts that everybody will be free at some time in the future,
but it does not imply that everybody will simultaneously be free. Thus I could
be free at time tl, you could be free at time t2, and so on. Now (3.30) implies
(3.31), but the converse does not hold.

Some Common Temporal Properties*

The temporal component of £ORA was introduced in order to allow us to

specify the dynamics of a system: how the state of an agent, and the environ­
ment itself, change over time. The aim in this section is to consider the types
of properties we find ourselves specifying using this temporal component. As
it turns out, the most important types of such properties are liveness and safety
properties.2
Informally, a safety property can be interpreted as saying that "something
bad won't happen." More formally, a safety property states that every finite
prefix of a path satisfies some requirement. For obvious reasons, safety prop­
erties are sometimes called invariance properties. The simplest kind of safety
property is global invariance, expressed by formulae with the following shape.

Dcp

A local invariance states that whenever cp holds, 'I/; must hold also. Such
invariances are specified by formulae with the following shape.

D(cp :::} '1/;)

Where a system terminates, partial correctness may be specified in terms of a

preconditioncp, which must hold initially, a postcondition '1/;, which must hold
on termination, and a condition X, which indicates when termination has been
reached.

2 The material in this section has been adapted from [49, p.1049--1054].
Introduction to .cORA 61

A mutual exclusion property is a global invariance of the form:

This formula states that at most one of the properties 'Pi E {'Pl,···, 'Pn}
should hold at any one time. The � notation arises if one imagines that
truth is valued at 1, falsity at 0; the above formula is read "at most one
of 'Pi E {'Pl, ..., 'Pn} is true." Any formula written in �-notation can be
expanded into an ordinary formula if required; the �-notation may therefore
be regarded as an abbreviation.
A liveness property is one that states that "something good will eventually
happen." The simplest liveness properties have the following shape.

Termination is an example of liveness. Basic termination properties are speci­

fied by formulae with the following shape.

This states that every path that initially satisfies the property 'P eventually sat­
isfies property X. Here X is the property that holds when a path has terminated.
Another useful liveness property is temporal implication.

O('P '* O'l/J)

Such formulae intuitively state that "every 'P is followed by a 'l/J." Respon­
siveness is a classic example of temporal implication. Suppose 'P represents
a "request," and 'l/J a "response." Then the above temporal implication would
state that every request is followed by a response. Another example of temporal
implication is total correctness, which states that a system initially satisfying
property 'P will eventually terminate (given by X), and on termination will sat­
isfy property 'l/J.

3.4 The Action Component of .cORA

The last component of .cORA is probably the hardest to understand. It is the

component that allows us to represent the actions that agents perform, and the
62 Chapter 3

Table 3.5
Operators for representing actions.

Formula Interpretation
(Happens 0) action expression 0 happens next
(Achvs 0 'P) action0 occurs, and achieves 'P
(Agts 0 g) group g is required to do action 0

effects of these actions. Before we do this, we need to say a little about how
actions fit into our temporal model.
The basic idea is that transitions between states are labeled with actions.
Intuitively, it is the perfonnance of an action by some agent that causes the
change in state. We assume that actions are atomic: only one action is ever
perfonned at any given time. However, we do allow for the same action
perfonned in the same state having multiple possible outcomes.
To illustrate these ideas, consider Figure 3. 2, in which the branching tem­
poral structure of Figure 3. 1 has been augmented by labeling state transitions
with the actions that cause them. For simplicity, we will assume in this exam­
ple that there is only one agent in the environment. From state to, the agent
perfonns action aI, which transfonns the environment to state tl. When the
environment is in state tl, the agent has a choice of two actions available to it.
It can either perfonn action a2 or it can perfonn a3. If it perfonns a2, then
either state t2 or state t3 can result. If instead it chooses to perfonn a3, then the
only possible outcome is state t4. From state t2, the agent has no choice about
what action to perfonn - it must perfonn a4, and either state t5 or state t6 will
result. From state t3, the agent can only perfonn action a5, and only state f7
can result. Finally, when the environment is in state t4, the agent has a choice
of perfonning either a6 or a3, with t8 or t9 resulting respectively.
.cORA provides a number of connectives that allow us to express the
properties of actions; these are summarized in Table 3. 5. In order to express the
fact that some action happens, we have a modal connective "Happens." This
connective is similar to 0, <>, and D in that it expresses a property of a path.
This operator takes a single argument, an action expression, and expresses the
fact that this action expression is the first thing that happens on the path. For
example, the fonnula

is true when evaluated on the path (tl, t2, t5, . . . ) , for the obvious reason that
a2 is the first action that happens on this path.
Introduction to .cORA 63

t5

t2

t6
a4

a2

t8
a3

a6

t9
a3

Figure 3.2
A branching temporal structure including actions that cause state transitions.

Table 3.6
Constructors for action expressions.

Expression Interpretation
Q;Q' a followed by a'
l
a a' either a or a'
a* a repeated more than once
cp? cp is satisfied

To be precise, the symbol a2 in the formula (Happens (2 ) is actually a

term of the language .cORA that stands for the actual action a2 . .cORA
allows us to combine terms that stand for actions to make action expressions
- see Table 3.6. The constructors for action expressions are very similar to
the constructs that one finds in imperative programming languages such as
'
PASCAL. Thus we can combine two action terms a and a to make an action
' '
expression a; a , which means a followed by a . This sequential composition
constructor behaves exactly like the semicolon in most imperative languages.
To make this concrete, consider again the path (t1' t2, t5, . . . ). Then the formula
64 Chapter 3

is true on this path, since a2 is the action that transforms t1 to t2, and a4
transforms t2 to t5.
The non-deterministic choice constructor, "I," simply means that one of its
arguments occurred. Thus the formula

is true on path ( t1' t2, t5, ...) , since one of a2 or a5 occurs on the first transi­
tion. So (Happens a l a') means that either a happens or a' happens.
The iteration constructor, "*," means that its argument occurs zero or
more times in succession. With respect to Figure 3.2, consider the path
( t1' t2, t5, ...) . The formula

(Happens a 2*)
is trivially true on this path, since a2 transforms t1 to t2. On the path
( t1' t4, t9, ...) , the action a3 occurs twice in succession, and so the formula

is satisfied on this path.

In all programming languages, there is some notion of "test," and .cORA
also provides for this. The test constructor, "7," allows us to build action
expressions that depend on the truth or falsity of formulae. To illustrate how
this works, consider the following formula.

(Happens q7) (3.32)

Formula (3.32) is true when evaluated on path ( t1' t2, t5, ...) , since the formula
q is true on the first state of this path, t1. However, (3.32) is false when
evaluated on the path ( t3, t7 , ...) , since q is not true in t3.It may not be obvious
why we have the "test" construct. To see why, consider the following formula.

(Happens ('1'7; a) I (""'1'7; a'))

This formula says that either ('1'7; a) will happen or (""'1'7; a') will happen.
Thus if 'I' is true then a will happen, else if ""'1' is true then a' will happen. So
the action expression ('1'7; a) I (""'1'7; a') captures the sense of a statement
'
if 'I' then a else a .

Similarly, consider the programming language statement

while 'I' do a
Introduction to CORA 65

We can capture the sense of this in the following action expression.

[(cp?; a) I (--, cp?)]*

Similarly, the programming language statement

repeat a until 'I'

can be represented by the following action expression.

a; [(--, cp?; a) I cp?]*

Finally, actions are usually performed by agents in order to bring about

certain states of affairs. We have a modal connective

(Achvs a '1')

which represents the fact that:

• the action expression a is possible, in that a occurs on some path originating

from the current state;
• after a is performed, 'I' is guaranteed to be true.

In order to allow us to designate the agents required to perform a sequence of

actions, we have an operator "Agts." This operator takes two arguments: the
first is an action expression, the second is a term denoting a set of agents. Thus
(Agts a g) means that the group denoted by g are precisely the agents required
to perform the actions in the expression a.
Before leaving this section, it is worth commenting on the issue of quan­
tifYing over action expressions. There is really only one rule that needs to be
remembered: We can quantify over sequences of actions, but we are not al­
lowed to quantify over arbitrary action expressions. To see what is meant by
this, consider the following formula.

3a· (Happens a)

In this path formula, the existentially quantified variable a is a term that

denotes a sequence of actions. It is not a term that denotes an arbitrary action
expression - such terms do not exist in .cORA. This restriction has a number
of implications. For example, it means that the following type of construction
is not legal in .cORA.

3'1' . (Happenscp?)
66 Chapter 3

(If we allowed this type of construction, then we would implicitly be quan­

tifying over formulae of CORA, which would lead to a number of serious
complications. )

3.5 Groups of Agents in CORA

CORA contains terms that denote groups of agents, and gives us a simple
but powerful set-theoretic apparatus to represent the properties of such groups.
The basic construct, with the obvious set-theoretic interpretation, is as follows.

(i E g)

This state formula asserts that the agent denoted by i is a member of the group
denoted by g.Here g is a term that stands for a (non-empty) set of agents.
As well as simply asserting that one agent is a member of a group, we have
the expected set-theoretic constructions (�, C) for expressing subset and other
relations between groups of agents.

3.6 Notes and Further Reading

CORA both draws upon, and extends, a number of previous formalisms.

In this section, I briefly consider the relationships between the formalism
presented in this chapter and other formalisms developed in the literature.
Syntactically, the logic resembles that of Cohen and Levesque [35]. Like
their logic, it has modalities for representing beliefs and goals (desires), and
a mechanism for describing the structure and occurrence of complex actions,
with action constructors loosely based on those found in dynamic logic [93].
However, it extends their language in several respects. First, it contains terms
denoting groups of agents, and provides a set-theoretic mechanism for relating
agents and groups at the object-level. In addition, our language takes intentions
as primitives - Cohen-Levesque argued that intentions could be reduced to
beliefs and goals. A more significant difference is that the new language con­
tains path quantifiers, for quantifying over possible future courses of events; in
this respect, the language is similar to those of [185, 245, 216]. Semantically,
CORA is very different to the logic of Cohen and Levesque. The most signifi­
cant point of difference is that worlds in the Cohen and Levesque language are
linear structures whereas worlds in CORA are themselves branching temporal
structures.
Introduction to CORA 67

A number of other logics for representing the properties of rational agents

have been proposed in the literature. Among the most mature of these is
the KARO framework of Meyer et al. [153, 137, 99]. KARO is perhaps best
viewed as a combination of dynamic logic, doxastic (belief) logic, and a
commitment logic. Meyer et al. focus in particular on the dynamics of mental
states: how actions can change an agents beliefs, desires, and so on. CORA
has much in common with KARO, the main difference being that CORA is
based on temporal logic, and builds a kind of dynamic logic on top of this
basic framework. In contrast, KARO builds on top of dynamic logic, and does
not have explicit temporal modalities. Semantically, KARO is rather complex
- even more so than CORA. See [153, pp.32-33] for a discussion on the
relationship of KARO to BDI logic.
Another well-established formal framework for reasoning about agents is
that developed by Singh [216]. Singh adopts a CTL *-based branching time
framework as his starting point, in which the basic structures, as in CORA,
are branching time labeled transition systems. However, Singh's formalism
is unusual in that it allows for concurrent actions to be performed; another
unusual property is that his formalism is based on continuous (as opposed
to discrete) time. Such a temporal framework is much richer than that which
underpins CORA, although one could argue that this richness comes at the
cost of complexity - the coherence constraints required to hold the technical
framework together are quite elaborate [216, pp.28-37]. Using his basic model,
Singh defines belief, intention, and several varieties of ability, and proves
completeness for some subsets of his logic.
4 CORA Defined

It is now time to give a complete formal definition of CORA. As I pointed out

in previous chapters, CORA is rather baroque from a logical point of view.
It extends full first-order branching time temporal logic with the addition of
modalities for referring to the beliefs, desires, and intentions of agents, as well
as with a dynamic logic-style apparatus for reasoning about actions.
Although the semantics of CORA are based upon techniques that are by
now standard in the modal and temporal logic communities, the combination
of so many logical features in the same language makes the formal definition
of CORA somewhat complex. As a result, this chapter contains some long
and involved definitions. You will not need to understand all these definitions
to make sense of most of the book, but if you are interested in following the
various proofs that subsequently appear, then some understanding is essential.
If you do want to understand the technical detail, then it may be worth reading
appendix B before attempting this chapter.
The logic is presented in four parts:

• In section 4.1, I formally define the syntax of the logic.

• In section 4.2, I present the semantics of the logic.
• In section 4.3 I present some derived connectives -widely used, useful
constructs that are not primitives, but are defined in terms of primitives.
• In section 4.4, I establish some properties of the logic.

4.1 Syntax

CORA is a many-sorted logic, which permits quantification over various types

of individuals: agents, (sequences of) actions, sets of agents (groups), and other
individuals in the world. All of these sorts must have a corresponding set of
terms (variables and individual constants) in the alphabet of the language.
In addition, CORA contains the belief, desire, and intention modalities, the
temporal connectives, the usual apparatus of first-order logic, and finally,
operators for talking about membership in groups and the agents of an action.
The alphabet of CORA thus contains the following symbols:

1. A denumerable set Pred of predicate symbols;

2. A denumerable set Const of constant symbols, the union of the following
70 Chapter 4

mutually disjoint sets:

• ConstAg - constants that stand for agents;
• ConstAc - constants that stand for action sequences;
• ConstGr - constants that stand for sets of agents (groups); and
• Constu - constants that stand for other individuals.
3. A denumerable set Var of variable symbols, the union of the following
mutually disjoint sets:
• VarAg - variables that denote agents;
• VarAc - variables that denote action sequences;
• VarGr - variables that denote sets of agents (groups); and
• Varu - variables that denote other individuals.
4. The modal connectives:
• "true" - a logical constant for truth;
• "Bel" - the belief modality;
• "Des" - the desire modality;
• "Int" - the intention modality;
• "A" - the universal path quantifier;
• "U " - the binary temporal "until" connective; and
• "0" - the unary temporal "next" connective.
5. The additional operators:
•
"E" - group membership; and
• "A gts " - the agents of an action.
6. The action constructors: "; ", "I", "*", and "?".
7. The classical connectives:
• "V" - "or" and "...," - "not".
8. The universal quantifier: "\;/".
9. The punctuation symbols " )", "( ", and ".".

Associated with each predicate symbol is a natural number called its arity,
given by the arity function.

arity : Pred ---+ IN

Predicates of arity 0 are known as proposition symbols.

CORA Defined 71

A sort is eitherAg, Ac, Gr, or U. If a is a sort, then the set Term,,, of terms
a/ sort a, is defined as follows.

Termu = Varu U Constu

The set Term, of all tenns, is defined by

Term = TermAg U TermAc U TermGr U Termu

We use 7 (with decorations: 7',71,...) to stand for members of Term. To

indicate that a particular tenn 7 is of type a, we subscript it as follows: 7u.
The syntax of CORA is then defined by the grammar in Figure 4.1 (it
is assumed that predicate symbols are applied to the appropriate number of
arguments).
Notice the use of parentheses in operators such as Bel, in order to make
explicit the binding between a modality and its arguments. To make the in­
tended interpretation of a fonnula clear, I also use parentheses and other brack­
ets where necessary. Finally, I assume the "standard" rules of binding between
logical connectives [52, pp.39--44].

4.2 Semantics

In general, syntax is easy but semantics is hard - CORA is no exception to

this rule. The semantics of CORA generalize those of first-order branching
time temporal logic. Although the basic underlying structures to CORA are
essentially labeled transition systems, which have been studied in theoretical
computer science for decades, the way these structures are used in CORA is
non-standard. I will therefore spend some time introducing and motivating the
use of these structures, before moving on to the fonnal presentation.

The Basic Components: Agents, Actions, and Time

CORA allows us to represent the properties of a system that may evolve in

different ways, depending upon the actions perfonned by agents within it. We
let DAg be the set of all agents, and let DAc be the set of all actions these agents
may perfonn.
As noted in chapter 3, the model of time that underpins CORA is:

• discrete;
• bounded in the past (there is a "start of time");
72 Chapter 4

(ag-term) .. - any element of TermAg /* agent terms */

(ac-term) .. - any element of TermAc /* action terms */
(gr-term) .. - any element of TermGr /* group terms */
(ac-exp) .. - (ac-term)
(ac-exp) ; (ac-exp) /* sequential composition */
(ac-exp) "I" (ac-exp) /* non-deterministic choice */
(state-Jmla) ? /* test actions */
(ac-exp)* /* iteration */
(term) .. - any element of Term /* arbitrary terms */
(pred-sym) .. - any element of Pred /* predicate symbols */
(var) .. - any element of Var /* variables */
(state-Jmla) .. - true /* truth constant */
(pred-sym) ((term), . . . , (term)) /* predicates */
(Bel (ag-term) (state-Jmla)) /* belief formulae */
(Des (ag-term) (state-Jmla)) /* desire formulae */
(Int (ag-term) (state-Jmla)) /* intention formulae */
(A gts (ac-exp) (gr-term)) /* agents of an action */
((term) (term))
= /* equality */
((ag-term) E (gr-term)) /* group membership */
A(path-Jmla) /* path quantifier */
,(state-1mla) /* negation */
(state-Jmla) V (state-1mla) /* disjunction */
V(var) . (state-Jmla) /* quantification */
(path-Jmla) .. - (Happens (ac-exp)) /* action happens */
(state-Jmla) /* state formulae */
(path-Jmla) U (path-Jmla) /* until */
o (path-Jmla) /* next */
,(path-Jmla) /* negation */
(path-Jmla) V (path-Jmla) /* disjunction */
1 V(var) . (path-Jmla) /* quantification */

Figure 4.1
The syntax of CORA.

• unbounded in the future (there is no "end of time");

• linear in the past (there is only one past history); and
• branching in the future (the course of future events is yet to be determined).

In order to represent such a model, we use a branching temporal structure. A

branching temporal structure is a total, backwards-linear directed graph over
a set of T of time points. Note that a relation Ris total if every node in Rhas
at least one successor. Thus the binary relation Rover T is total if it satisfies
the following condition.

\:It· t E T => 3t' . [t' E T and ( t,t' ) E R]

CORA Defined 73

We let R � TxT be the branching temporal structure that encodes all ways in
which the system can evolve. Arcs in R will correspond to the performance of
atomic actions by agents within the system, and we therefore label these arcs
with the actions that they correspond to via the functionAct.

Act: R ---+ DAc

Each action is associated with a single agent, given by the function Agt.

Agt : DAc ---+ DAg

Possible Worlds

The state of an agent is defined by its beliefs, desires, and intentions. The
semantics of beliefs, desires, and intentions are given using possible worlds,
or Kripke semantics [31]. Thus an agent's beliefs in any given situation are
characterized by a set of situations - those that are consistent with the agent's
beliefs. An agent is then said to believe r.p if r.p is true in all these possible
situations. We refer to this set of "belief alternatives" as belief-accessible
situations. Similarly, an agent's desires in any given situation are characterized
as a set of situations - those that are compatible with the agent's desires. As
might be expected, we refer to these as desire-accessible situations. Finally, an
agent's intentions in a given situation are characterized by a set of intention­
accessible situations, each of which represents a state of affairs compatible
with the agent's intentions.
At this point, the semantics of our framework appear to be quite conven­
tional. However, worlds in CORA are not instantaneous states but are them­
selves branching time structures. The intuition is that such structures represent
an agent's uncertainty not only about how the world actually is, but also about
how it will evolve.
Formally, a world is a pair (T', R'), where T' � T is a non-empty set of
time points, and R' � R is a branching time structure on T'. Let W be the set
of all worlds over T. Formally, W is defined as follows.

W = {(T', R') I T' � T, R' � R, and T' = ( dom R' U ran R'H

If w E W is a world, then we write Tw for the set of time points in w, and Rw

for the branching time relation in w.
74 Chapter 4

A pair (w, t), wherewE W and t E Tw , is known as a situation. IfwE W,

then the set of all situations inwis denoted by Sw .

Sw = { (w, t) I wE W and t E Tw }

In order to characterize the beliefs of each agent, we use a function B, which

assigns to every agent a relation over situations.

B : DAg -+ p( W x T x W ).

I will refer to functions B as belief accessibility relations, even though, strictly

speaking, they are functions that assign belief accessibility relations to agents.
In order to simplify what follows, I write B�(i) to denote the set of worlds
accessible to agent i from situation (w, t). Formally, B�(i) is defined as follows.

B7(i) = {w' I (w, t,w' ) E B(in

Belief accessibility relations are required to satisfy several properties. The first
is that if a world w' is accessible to an agent from situation (w, t), then t is
required to be a time point in bothwandw'. Formally, this requirement (which
we call world/time point compatibility) is captured as follows.

Ifw' E B7(i) then t E wand t E w'. (4.1)

We also require that the relation that the function B assigns to every agent is
serial, transitive, and Euclidean. B is said to be:
• serial if for all situations (w, t), there is some worldw' such thatw' E B�(i);
• transitive ifw' E Br(i) andw" E B�' (i) impliesw" E B�(i); and
• Euclidean ifw' E B�(i) andw" E B�(i) impliesw' E B�" (i).

These requirements ensure that the logic of belief corresponds to the well­
known modal logic system KD45 [31] we will see a proof of this later.
-

To give a semantics to desires and intentions, we proceed in a similar

fashion. Thus an agent's desires are given by a function D.

D : DAg -+ p( W x T x W)

An agent's intentions are given by a function I.

I: DAg -+ p( W x T x W)

Both D and I are assumed to assign agents serial relations. This ensures that
the desire and intention modalities have a logic of KD [31]; we also require that
CORA Defined 75

both D and I satisfy the world/time point compatibility requirement property,

analogous to (4.1) above. As with belief accessibility relations, we write D� (i)
and I(' (i) to denote the desire- and intention-accessible worlds for agent i from
situation (w,t).
We now present some technical apparatus for manipulating branching time
structures. Let w E W be a world. Then a path through w is a sequence
(to,t1' . . .) of time points, such that for all u E IN, we have (tu,tu+d E Rw .
Let paths(w) denote the set of paths through w. In the literature on branching
temporal logics, paths are sometimes referred to as fullpaths [50]. If P is a path
and u E IN, then p(u) denotes the u + 1 'th element of p. Thus p(O) is the first
time point in p, p( 1) is the second, and so on. If p is a path and u E IN, then
the path obtained from p by removing its first u time points is denoted by p( ).
u

Domains

We now need to identify the objects that we can refer to using CORA. The
language will contain terms that stand for these objects, and in particular, these
are the objects over which we can quantify. These objects represent the domain
of CORA the universe ofdiscourse.
-

The domain of CORA contains agents, given by the set DAg, as well as
actions (given by the set DAc). We will also allow quantification over sequences
of actions, i.e., the set DAc' groups of agents, given by the set DGr (where
DGr � p(DAg)), and other objects (pints of beer, chairs, pieces of string, etc.)
given by a set Du.
Putting these components together, a domain (of interpretation), D, is a
structure:

where:

• DAg = {I,. . . ,n} is a non-empty set of agents;

• DAc { a, a', . . . } is a non-empty set of actions;
=

• DGr = p(DAg) \ {0} is a set of non-empty subsets of DAg, i.e., the set of

agent groups over DAg;

• Du is a non-empty set of other individuals.

If D = (DAg, DAc, DGr, Du) is a domain, then we denote by lJ the set

lJ = DAg U DAc U DGr U Du.

76 Chapter 4

(Notice thatD includes the set of sequences over DAc, rather than simply the
set DAc.) If D is a domain and u E IN, then byDU we mean the set of u-tuples
overD.

DU=Dx···xD
'-v-"
u times
Denotation

In order to give a meaning to formulae of CORA, we need various functions

that associate symbols of the language with objects in the domain. The first of
these is an interpretation/or predicates. This is a function that assigns to every
predicate at every time point a set of tuples that represent the extension of the
predicate in that time point. Formally, an interpretation for predicates, <1>, is a
function

<I>: Predx T-t p( U DU).

uEIN

The function <I> is required to preserve arity. In other words, if the function
assigns a predicate symbol P a set of u-tuples, then the arity of P should be u.
Formally, this condition can be stated as follows.

If arity(P)= u then <I>(P,t) � DU for all t E T.

An interpretation/or constants, C, is a function that takes a constant and a

time point and returns the denotation of that constant in that time point.

C: Canstx T-tD

This function is required to preserve sorts, in that if a E Constu and t E

T, then C(a,t) E Du . Note that constants do not necessarily have a fixed
interpretation: a constant a that denotes d E D in time t E T may denote
d' E D at time t' E T, where d -I- d'. However, in the remainder of the book, I
will assume that constants have afixed interpretation unless otherwise stated,
as this clarifies many results.
A variable assignment, V, associates variables with elements of the do­
main.

V: Var-tD

As with interpretations for constants, if x E Varu, then V(x) E Du (i.e.,

variable assignments preserve sorts). In contrast to constants, the interpretation
CORA Defined 77

of variables is fixed across all time points.

We now introduce a derived function [ . . .]v,c, which gives the denotation
of an arbitrary term. If V is a variable assignment and C is an interpretation for
constants, then by [ . . .]v,c, we mean the function

[ . . .]v,c: Termx T --t jj

that interprets arbitrary terms relative to V and C, as follows:

[T, t]v,C =
A { C(T, t) ifTE Const
V(T) otherwise.
Since V, C, and the time point in which we are evaluating formulae will always
be clear from context, reference to them will usually be suppressed: we simply
write [T] .

Models and Satisfaction Relations

A model for CORA is a structure

M = (T,R, W,D,Act,Agt,B,V, I, C, <1»

where:

• T is the set of all time points;

• R � TxT is a total, backwards-linear branching time relation over T;
• W is a set of worlds over T;
• D = (DAg, DAc, DGr, Du) is a domain;
• Act: R --t DAc associates an action with every arc in R;
• Agt: DAc --t DAg associates an agent with every action;
• B: DAg --t p(Wx T x W) is a beliefaccessibility relation;
• V: DAg --t p(Wx T x W) is a desire accessibility relation;

• I: DAg --t p(Wx T x W) is an intention accessibility relation;

• C: Constx T --t jj interprets constants; and
• <1>: Predx T --t P(UuElN jjU) interprets predicates.

The semantics of the language are defined in two parts, for path formulae
and state formulae respectively. The semantics of path formulae are given
via the path formula satisfaction relation, "FP", which holds between path
formulae interpretations and path formulae. A path formula interpretation is
a structure (M, V, w, p) , whereM is a model, V is a variable assignment, w is
78 Chapter 4

(M, V, w, p) FP cp iff (M, V, w, p(O)) FS cp (where cp is a state formula)

(M, V, w, p) FP -,cp iff (M, V, w, p) �p cp
(M, V, w, p) FP cP v 1/1 iff (M, V, w, p) FP cp or (M, V, w, p) FP 1/1
(M, V, w, p) FP 'tx· cP iff (M, V t {x >-+ d}, w, p) FP cP
for all d E !J such that x and d are of the same sort
(M, V, w, p) FP cpU 1/1 iff:lu E lN such that (M, V, w, p(u)) FP 1/1 and
'tv E lN, if (0 :'S v < u), then (M, V, w, p(v)) FP cP
(M, V, w, p) FP Ocp iff(M, V, w, p(1)) FP cP
(M, V, w, p) FP (Happens a) iff :lu E lN such that occurs(a, p, 0, u)

Figure 4.2
Rules defining the semantics of path formulae. Note that the rule for Happens makes use of the
occurs predicate, which is defined in the text.

a world in M, and p is a path through w. The rules defining the path formula
satisfaction relation are given in Figure 4.2. If (M, V, w,p) FP <p, then we say
(M, V, w,p) satisfies <p, or equivalently, that <p is true in (M, V, w,p).
The semantics of state formulae are given via the state formula satisfaction
relation, "Fs". This relation holds between state formulae interpretations
and state formulae. State formulae interpretations are structures of the form
(M, V, w,t), where M is a model, V is a variable assignment, w is a world
in M, and t E Tw is a time point in w. The rules defining this relation are
given in Figure 4.3. As with path formulae, if (M, V, w,t) FS <p, then we say
(M, V, w,t) satisfies <p, or <p is true in (M, V, w,t).

{
The rule for the Agts connective makes use of a function Agts, which
extends the Agt function to arbitrary action expressions.
{Agt(ad,...,Agt(ak)} where [a,t] = al,...,ak
( ) Agts(al,t)UAgts(a2,t) where a = al; a2 or a = al I a2
Agts a,t =
A

Agts(al,t) where a = al*

o where a is of the form <p?
The first case in this inductive definition deals with the sequences of actions.
The second case deals with sequential composition and non-deterministic
choice. The third case deals with iteration, and the final case with test actions.

The Semantics of Action Expressions

The semantic rules for path formulae make use of an important auxiliary
definition: occurs. This definition is used to define the Happens operator.
We write occurs(a,p,u,v ) to indicate that action a occurs between "times"
CORA Defined 79

(M, V, w, t) FS true
(M, V, w, t) FsP(Tl,... ,Tn) iff (h], ... , [Tn]) E <l>(P,t)
(M, V, w, t) FS -''P iff(M, V, w, t) ftos 'P
(M, V, w, t) FS 'P V 1/1 iff(M, V, w, t) FS 'P or (M, V, w, t) FS 1/1
(M, V, w, t) FS \Ix· 'P iff(M, V t {x t-+ d}, w, t) FS 'P
for all d ElJ such that x and d are of the same sort
(M, V, w, t) FS (Bel i 'P) iff\lw' E W, ifw' E B7([i]), then (M, V, w', t) FS 'P
(M, V, w, t) FS (Goal i 'P) iff\lw' E W, ifw' E 1J7([i]), then (M, V, w', t) FS 'P
(M, V, w, t) FS (Int i 1/1) iff\lw' E W, ifw' E:r;"([i]), then (M, V, w', t) FS 'P
(M, V, w, t) FS (Agts a g) iffAgts( a, t) = [g]
(M, V, w, t) FS (T T')
= iff[T] =[T']
(M, V, w, t) FS (i E g) iff[i] E [g]
(M, V, w, t) FS A'P iff\lp Epaths(w), ifp(O) = t, then (M, V, w, p) F1' 'P

Figure 4.3
Rules defining the semantics of state formulae.

u,V E IN on the path p. Note that occurs is a meta-language predicate. It is

a predicate that we use in our presentation of .cORA, rather than a predicate
that would be used in a formula of .cORA itself.
The occurs predicate is best illustrated by an example. Suppose we have a
path p through a world w, in which the arcs are labeled as follows.

Then:

• occurs ( ao,p, 0, 1) holds, as the arc (P( O) ,p ( l )) , i.e., ( to,t d , is labeled with

0.0;
• occurs ( ao; 0.1,p, 0,2) holds, as occurs ( ao,p, 0, 1) holds (this was the first

example), and occurs( a1,p, 1,2) holds, since the arc (P( 1) ,p(2)) , that is,
( t1' t2) , is labeled with the action 0.1;
• occurs ( ao; (0.1* ) ,p, 0,4) holds, as occurs ( ao,p, 0, 1) holds (this was the

first example), and occurs( a1*,p, 1,4) holds. To see that occurs( a1*,p, 1,4)
holds, note that occurs( a1,p, 1,2) holds and occurs( a1*,p,2,4) holds. To
see that occurs( a1*,p,2,4) holds, observe that occurs( a1,p,2, 3) holds and
occurs( a1*,p,3, 4) holds since occurs( a1,p,3, 4) holds.
80 Chapter 4

I leave it to the reader to use similar reasoning to show that the following
also hold.

occurs(a1*; a2,p, 1, 5)

Formally, the occurs predicate is defined inductively by five rules: one for each
of the action expression constructors, and one for the execution of primitive
actions. The first rule represents the base case, where a primitive action is
executed.
occurs(a,p,u,v ) iffv = u + k and [a,p(u)] = 0'.1,...,ak and
Act(p(u),p(u + 1)) = 0'.1,...,Act(p(u + k - l ),p(u + k)) = ak (4.2)
(where a E TermAc )
Thus primitive action a occurs on path p between u and u + 1 if the arc from
u to u + 1 is labeled with a.
The second rule captures the semantics of sequential composition. The
action expression a; a' will occur between times u and v iff there is some time
point n between u and v such that a is executed between u and n, and a' is
executed between n and v.

occurs(a; a' ,p,u,v ) iff:3n E {u,... , v} such that (4.3)

occurs(a,p,u,n) and occurs(a' ,p,n,v )
The semantics of non-deterministic choice are even simpler. The action ex­
pression a I a' will be executed between times u and v iff either a or a' is
executed between those times.

occurs(a I a' ,p,u,v ) iff occurs(a,p,u,v ) or occurs(a' ,p,u,v ) (4.4)

The semantics of iteration rely upon the fact that executing 0'.* is the same
as either: (i) doing nothing, or (ii) executing a once and then executing 0'.*.
This leads to the followingfixed point rule, where the right-hand side is defined
in terms of the left-hand side.

occurs(a*,p,u,v ) iff u = v or occurs(a; (a*),p, u,v ) (4.5)

Finally, we have a rule that defines the semantics of test actions. The idea is that
action expression 'P? occurs between times u and v on path p if 'P is satisfied on
p at time u. This rule recursively makes use of the definition of whether or not
a formula is satisfied; in practice, the interpretation against which the formula
CORA Defined 81

is being evaluated will always be clear from context.

occurs(rp?,p,u,v) iff (M, V,w,p(u)) FS rp (4.6)

4.3 Derived Connectives

In addition to the basic connectives defined above, it is useful to introduce

some derived constructs. These derived connectives do not add to the expres­
sive power of the language, but are intended to make formulae more concise
and readable. We say they are "derived" because they are all defined in terms
of existing constructs of the language.
We begin by assuming that the remaining connectives of classical logic
(i.e., "1\" -"and," " => " "if. . . then. . . ," and "¢:?"
- - "if, and only if') have
been defined as normal, in terms of "...," (not) and "V" (or).
...,(...,rp V ...,'lj;)
(...,rp V 'lj;)
((rp => 'lj;) 1\ ('lj; => rp))
Similarly, we assume that the existential quantifier, "3 ", has been defined as
the dual of "\I".

3x . rp � ..., \lx . ""rp

Derived Temporal Operators

Next, we introduce the existential path quantifier, "E", which is defined as the
dual of the universal path quantifier "A". Thus a formula Erp is interpreted as
"on some path, rp," or "optionally, rp":

It is also convenient to introduce further temporal connectives. The unary con­

nective "0" means "sometimes." Thus the path formula Orp will be satisfied
on some path if rp is satisfied at some point along the path. The unary " D"
connective means "now, and always." Thus D rp will be satisfied on some path
if rp is satisfied at all points along the path. We also have a weak version of the
"u" connective: rp W 'lj; is read "rp unless 'lj;."

� trueU 'P
"" O...,rp
(rpU'lj;) V D rp .
82 Chapter 4

Thus cp W 'lj; means that either: (i) cp is satisfied until 'lj; is satisfied; or else (ii) cp
is always satisfied. It is said to be weak because it does not require that 'lj; be
eventually satisfied.

Derived Operators for Reasoning about Groups

CORA provides us with the ability to use simple set theory to relate the
properties of agents and groups of agents. The operators � and c relate groups
together, and have the obvious set-theoretic interpretation; (Singleton g i)
means g is a singleton group with i as the only member; (Singleton g) simply
means g is a singleton.
�
(g � g') Vi· (i E g) :::} (i E g')
�
(g C g') (g � g') A -,(g = g')
�
(Singleton g i) Vj· (j E g) :::} (j = i)
�
(Singleton g) :3i· (Singleton g i)
Derived Operators for Reasoning about Actions

We will abuse notation a little by writing (Agt a i) to mean that i is the only
agent required to perform action a.

(Agt a i) � Vg· (Agts a g) :::} (Singleton g i)

We will frequently need to formalize statements along the lines of "a is

a good plan for cp." In order to do this, we use some definitions that closely
resemble dynamic logic [93]. The first of these is the derived operator Poss.
The intuition is that (Poss a cp) means that the action a occurs on at least one
possible future, and after a is executed, cp is true.

(Poss a cp) � E ( Happens a; cp?)

If you are familiar with dynamic logic, you will recognize that (Poss a cp)
corresponds approximately to (a)cp. Similarly, we can capture the dynamic
logic [a]cp through the Nee operator.

( Nee a cp) � A ( ( Happens a) :::} ( Happens a; cp?))

Thus ( Nee a cp) simply means that on all paths where a occurs, cp is true
immediately after a. However, ( Nee a cp) does not imply that a occurs on any
path - it simply says that if it does, then cp follows.
The (Aehvs a cp) operator is defined as a combination of Nee and Poss.
(Aehvs a cp) means that a does indeed happen on some path, and after a
CORA Defined 83

happens, <p is guaranteed to be true.

(Aehvs Q <p ) � (Poss Q <p ) A (Nee Q <p )

4.4 Some Properties of CORA

After introducing a new logic by means of its syntax and semantics, it is usual
to illustrate the properties of the logic by means of a Hilbert-style axiom
system. However, no complete axiomatization is currently known for eTL * ,
the branching temporal logic that underpins CORA. 1 Completeness proofs for
temporal modal logics, even in the propositional case, are relatively few and far
between [54, p.307]. For these reasons, I stay primarily at the semantic level,
dealing with valid formulae, rather than attempt a traditional axiomatization.
We use conventional definitions of validity and satisfiability, complicated
slightly by the fact that we have essentially two languages, for path and state
formulae.
Formally, a path formula<p is said to be satisjiableifffor some (M, V, w,p),
we have (M, V, w,p) FP <p; it is said to be valid iff for all (M, V, w,p), we
have (M, V, w,p) FP <po We indicate that <p is a valid path formula by writing

Similarly, the state formula <p is said to be satisfiable iff for some (M, V, w,t) ,
we have (M, V, w,t) FS <p; it is said to be valid iff for all (M, V, w,t) , we have
(M, V, w,t) FS <po We indicate that <p is a valid state formula by writing

FS <P
It is easy to see that state formula validity implies path formula validity.

LEMMA 4.1: If FS <p then FP <po

Proof State formulae are also path formulae, and the semantic rules for
evaluating them are identical. _

1 Proving completeness for an axiomatization of CTL * has been a major open problem in temporal
logic since the logic was first formulated in the mid-1980s. Stirling reported an axiomatization
in [227], though it subsequently transpired that the completeness proof contained a bug. Mark
Reynolds from Murdoch University in Australia reported a complete axiomatization in 1998,
though at the time of writing, the work was still under review. To find out more, I urge you to
contact Mark directly.
84 Chapter 4

This Lemma means that certain formulae will be valid by virtue of their shape,
irrespective of whether or not they are actually path formulae or state formulae.
For example, presented with a formula 'l/J that has the shapecp V -'cp, we do not
need to worry about whether 'l/J is a path formula or a state formula: whichever
it is, we know it will be valid. To indicate that a formula cp is valid in this
regard, we writeF cpo
We can establish a similar result with respect to satisfiability.

LEMMA 4.2: If cp is a satisfiable state formula, then cp is a satisfiable path

formula.

Since the semantics of CORA generalize those of classical logic, classical

modes of reasoning can be used in CORA.

THEOREM 4.1:

1. If cp is a substitution instance of a propositional logic tautology, thenF cpo

2. IfF cp => 'l/J andF cp, then F 'l/J.

Proof For (1), simply observe that the semantics of the propositional connec­
tives "-," and "V" are identical to classical logic. Similarly, for (2) - which
is simply modus ponens - the reasoning is identical to classical propositional
or first-order logic. •

CORA also has the properties one would expect of a sorted first-order logic.
We use the normal form cp(x ) to mean that the variable x isfree in cp, where
"free" has its standard interpretation. Additionally, cp[x /y] is used to denote the
formula obtained from cp by systematically replacing every occurrence of yin
cp by x (i.e., normal first-order substitution - see, e.g., [52, pp.104-106]).

THEOREM 4.2:

1. F Vx ·cp(x ) => cp[T/X] (whereTis a term that does not occur free incp, x and
Tare the same sort, andcp does not contain any modal or temporal connectives).
2. F (T = T).
3. F (T = T' ) => (cp => cp[T' /TD (where T' does not occur in the scope of
any modal or temporal operator in cp).

Proof (1) follows immediately from the semantics of universal quantifica-

CORA Defined 85

tion. (2) and (3) follow immediately from the semantics of equality. _

Next, we tum to the belief operator, Bel. This is essentially a normal modal
necessity connective, with semantics given via a serial, Euclidean, and transi­
tive accessibility relation [31]. Thus the logic of Bel corresponds to the well­
known normal modal system "weak-S5," or KD45. Although the proofs for
Theorem 4.3 are readily available in the literature, they are included here for
completeness; see the further reading section for references.

THEOREM 4.3:

1. FS (Bel i ('P:::} 'I/J)):::} ( (Bel i 'P):::} (Bel i 'I/J)).

2. F S (Bel i 'P):::} ..., (Bel ;""'P).
3. F S (Bel i'P):::} (Bel i (Bel i'P)).
4. F S ..., (Bel i 'P):::} (Bel ;"" (Bel i 'P)).
5. IfF S 'P thenF S (Bel i'P).

Proof For (1), suppose (M, V,w,t) F S (Bel i ('P:::} 'I/J)) and (M, V,w,t) F S
(Bel i 'P) for arbitrary (M, V,w,t). From the semantics of Bel, we know
that (M, V,w',t) F S 'P :::} 'I/J and (M, V,w',t) F S 'P for all w' such that
w' E B�([i]). Hence (M, V,w',t) F S 'I/J, and so (M, V,w,t) F S (Bel i 'I/J).
For (2), suppose not. Then for some (M, V,w,t) we have both (M, V,w,t) F S
(Bel i 'P) and (M, V,w,t) F S (Bel i ""'P). Since the belief accessibility
relation is serial, we know there will be at least one w' E Br([i]), and
from the semantics of Bel, we therefore know that (M, V,w',t) F S 'P and
(M, V,w',t) F S ""'P, hence (M, V,w',t) F S 'P and (M, V,w',t) �s 'P,
which is a contradiction. For (3), assume (M, V,w,t) F S (Bel i 'P) for ar­
bitrary (M, V,w,t). We need to show that (M, V,w,t) F S (Bel i (Bel i 'P)),
which by the semantics of Bel amounts to showing that (M, V,w',t) F S
(Bel i 'P) for all w' E Br([i]), which in tum amounts to showing that
(M, V,w",t) F S (Bel i 'P) for all w" E B�'([i]). But since the belief ac­
cessibility relation is transitive, if w' E Br([i]), and w" E B( ([i]), then
w" E Br([i]), and since (M, V,w,t) F S (Bel i 'P), then by the semantics
of Bel, we have (M, V,w",t) F S (Bel i 'P), and we are done. For (4), sup­
pose (M, V,w,t) F S ..., (Bel i 'P). Then for some w' E Br([i]), we have
(M, V,w',t) F S ""'P. Now, consider any world w" E Br([i]). The belief ac­
cessibility relation is Euclidean, and so since w' E Br([i]) and w" E Br([i])
we have w' E B( ([i]). Hence (M, V,w",t) F S ""'P, and since w" was ar-
86 Chapter 4

bitrary, we therefore have (M, V,w,t) F S (Bel i -,(Bel i cp)). For (5), as­
sume that F S cp. We need to show that (M, V,w,t) F S (Bel i cp) for all
(M, V,w,t). But since F S cp, then by definition cp is satisfied by all interpre­
tation structures, and in particular, (M, V,w',t) F S cp for all w' E Br([ i ]).
Hence (M, V,w,t) F S (Bel i cp). •

Note that (1) in Theorem 4.3 is usually known as the "K" axiom; (2) is known
as the "D" axiom; (3) as the "4" axiom, and (4) as the "5" axiom. Property
(5) is known as beliefnecessitation.
The Des and Int connectives have a logic that corresponds to the normal
modal system KD. (The proofs are identical to those of Theorem 4.3, and are
therefore omitted.)

THEOREM 4.4:

1. F S (Des i (cp=>'I/J))=>((Des i cp)=>(Des i 'I/J)).

2. F S (Des i cp)=> -,(Des i -,cp).
3. IfF S cp thenF s (Des i cp).

THEOREM 4.5:

1. F S (Int i (cp=>'I/J))=>((Int i cp)=>(Int i 'I/J)).

2. F S (Int i cp)=> -,(Int i -,cp).
3. IfF S cp thenF s (Int i cp).

Turning to the temporal connectives in CORA, state formulae have a number

of properties with respect to path quantifiers.

THEOREM 4.6: If cp is a state formula, then: F S cp ¢:} Ecp ¢:} Acp.

Proof Assume cp is a state formula. Then we need to prove three cases.

• cp=> Ecp
Assume (M, V,w,t) F S cp for arbitrary (M, V,w,t). Then for any path
p E paths(w) such that p(O) = t, we know that (M, V,w,p) F P cp, so
(M, V,w,t) F S Ecp.
• Ecp=> Acp

Assume (M, V,w,t) F S Ecp for arbitrary (M, V,w,t). Then by the semantics
of E, we know that (M, V,w,p) F P cp for some p E paths(w) such that
CORA Defined 87

p(O) = t. From the semantics of state formulae interpreted on paths, we

thus have (M, V,w,t) FS <p, and so for any path p ' E paths(w) such that
p(O) = t, we have (M, V,W,p ') FP <p and so from the semantics of A, we
know (M, V,w,t) FS A<p.
• A<p::::} <p

Assume (M, V,w,t) FS A<p for arbitrary (M, V,w,t). Then by the semantics
of A, we know that (M, V,w,p) FP <p for allp E paths(w) such thatp(O) = t.
Since the branching time relation is total, we know there is at least one such
path, and so we have (M, V,w,t) FS <po

This completes the proof. _

The universal path quantifier A also behaves rather like a normal modal
necessity connective with logic S5.

THEOREM 4.7:

1. FS A(<p::::} 'l/J)::::} ((A<p)::::} (A'l/J)).

2. FS A<p::::} -,(A-,<p).
3. FS A<p::::} <p (for state formulae <p).
4. FS A<p::::} AA<p.
5. FS -,A<p::::} A-,A<p.
6. If FP <p then FS A<p.

Proof For (1), assume (M, V,w,t) FS A(<p ::::} 'l/J) and (M, V,w,t) FS A<p
for arbitrary (M, V,w,t). Then (M, V,w,p) FP <p::::} 'l/J and (M, V,w,p) FP
<p for all p E paths(w) such that p(O) = t. Hence (M, V,w,p) FP 'l/J for
all p E paths(w) such that p(O) = t, and so (M, V,w,t) FS A('l/J) and we
are done. For (2), assume (M, V,w,t) FS A<p and (M, V,w,t) FS A-,<p for
arbitrary (M, V,w,t). Then from the semantics of A, we have (M, V,w,p) FP
<p and (M, V,w,p) FP -,<p for all p E paths(w) such that p(O) = t.
(Since the branching time relation is total, we know there will be at least
one such path.) Hence both (M, V,w,p) FP <p and (M, V,w,p) �p -'<p,
which is a contradiction. For (3), simply observe that this is one implication
of Theorem 4.6 (see above). For (4), assume (M, V,w,t) FS A<p for arbitrary
(M, V,w,t). We need to show that (M, V,w,t) FS AA<p. Since A<p is a state
formula, we know from Theorem 4.6 that FS A<p ¢:} AA<p, and so we have
(M, V,w,t) FS AA<p and we are done. For (5), assume (M, V,w,t) FS -,A<p.
88 Chapter 4

Then by definition, (M, V,w,t) FS E.<p. Hence from Theorem 4.6, we have
(M, V,w,t) FS A E.<p, and so (M, V,w,t) FS A.A<p. Finally, (6) is just a
version of necessitation, for which we have already seen proofs. _

The temporal logic component of CORA has been examined extensively in

the literature, and we will not discuss it further here - see [143, pp.228-234]
and [49, pp.lO03-1005].
The following theorem captures some properties of action expressions and
the Happens operator.

THEOREM 4.8:

1. F'P (Happens a; 0:') {:} (Happens a; (Happens a')7).

2. F'P (Happens a I 0:') {:} (Happens a) V (Happens a').
3. F'P (Happens <p7) {:} <p.
4. F'P (Happens a*).
5. F'P 3a· (Happens a).

Proof For (1), assume (M, V,w,p) F'P (Happens a; a') for arbitrary
(M, V,w,p). Then by the semantics of Happens, we have some u,v E IN
such that occurs(a,p, 0, v ) and occurs(a',p,u,v ) . Hence (M, V,w,p) F'P
(Happens a; (Happens 0:')7), and we are done. For (2), start by assuming
(M, V,w,p) F'P (Happens alo:'). Then by the semantics of Happens, we
must have either occurs(a,p, 0, u) for some u E IN or else occurs(0:',p, 0, u)
for some u E IN. Hence by the semantics of V and Happens, we have
(M, V,w,p) F'P (Happens a)V(Happens a'). For (3), assume (M, V,w,p) F'P
(Happens <p7). Hence we have (M, V,w,p(O)) F'P <p, and since <p must be
a state formula, we have (M, V,w,p(O)) FS <p and so (M, V,w,p) F'P <p.
Part (4) follows trivially from the semantics of iteration: any action happens
zero or more times. For (5), observe that time is a total branching time rela­
tion; hence there must be a path emerging from any situation, and the arcs that
make up this path will be labeled with actions. (Part (5) can also be Seen to be
implied by (4).) _

The fourth part of this Theorem may seem slightly strange; it makes sense
when one reads it as "every action occurs zero or more times."
CORA Defined 89

4.5 Notes and Further Reading

The semantics of the first-order logic component of CORA are relatively

standard; see Enderton's classic [52] or Genesereth and Nilsson's somewhat
more accessible [73] for similar presentations.
The 8DI component of CORA draws upon a long tradition of using modal
logics to represent mental states. laakko Hintikka is usually credited with
starting this work with the publication of his book Knowledge and Belief [96],
building on the work of Hendrik von Wright and others. Hintikka's legacy is
most clearly evident through the vast literature on epistemic logic [54, 152].
Robert Moore (building on the work of McCarthy and Hayes [151]) is often
credited with bringing the tools of modal logic to bear on reasoning about
agents in artificial intelligence [156, 157]. Since Moore's pioneering work, the
literature on the topic has grown enormously - see [249] for a survey.
The treatment of mental states in CORA owes a significant debt to the
intention logic developed by Phil Cohen and Hector Levesque [35]; this logic
has been widely used in the agents community for formalizing everything
from speech acts [36, 68] to social structure and dependence [27]. However,
CORA's greatest debt is obviously to Anand Rao and Michael Georgeff's
family of 8DI logics, developed in a sequence of papers in the first half of
the 1990s [186, 190, 187, 188, 189, 184, 79, 185]. In particular, CORA
adopts their semantics for beliefs, desires, and intentions, whereby worlds in
the possible worlds semantics are assumed to be themselves branching time
structures. A number of 8DI logics that include some apparatus for reasoning
about plans have appeared in the literature, including [190, 118, 87].
The temporal fragment of CORA is based on CTL *, the "full" branching
time temporal logic developed by Emerson and Halpern [50]; see appendix B
for notes on the origins of CTL *.
The action component of the logic is loosely based on dynamic logic [93,
122]. The actual treatment of action is based on that developed by Cohen and
Levesque [35].
The technical treatment of CORA is based on the logics in [245, 247,
248, 239]. Many of the ideas relating to use of groups directly as terms in the
language are developed in these papers.
Although there are many completeness results and proof methods for
modal logics with a single modality [66, 72, 30, 90] as well as temporal
logics [235, 82, 62, 135], there are comparatively few completeness results
or proof methods for logics that combine modal and temporal fragments.
90 Chapter 4

Preliminary proof methods for temporal logics of knowledge and belief are
described in [246, 47, 244]. With respect to 8DI logics, Rao reported proof
methods for linear-time 8DI logics in [184] and Rao and Georgeff reported
proof methods for branching time 8DI logic (where the branching temporal
component was restricted to CIL logic, rather than CIL*) in [185]. Klaus
Schild demonstrated that 8DI logic could be encoded in the J.l-calculus, and
established an exponential time upper bound on the satisfiability problem for
(cIL-based) propositional 8DI logic [200].
5 Properties of Rational Agents

In this chapter, we will see how CORA can be used to capture various
desirable (and undesirable) properties of rational agents. The chapter focuses in
particular on the possible interrelationships among the mental states of beliefs,
desires, and intentions. This study has two aspects:

• A philoso phi cal aspect, whereby we attempt to characterize the properties of

various types of rational agents in terms of formulae of CORA, and discuss the
extent to which these formulae adequately capture our intuitive understanding
of the concepts in question. For example, we might ask whether the formula
schema (Inti 'P) => (Bel i E<>'P) (ifi intends 'P, theni believes 'P is possible)
is an appropriate characterization of the relationship between intentions and
beliefs.
• A pu relylog i cal aspect, whereby we investigate the possible interrelation­
ships among beliefs, desires, and intentions from an abstract mathematical per­
spective. With such a perspective, we simply view CORA as a formal system
that has certain properties.We pay little or no attention to the question of what
these properties mightmean. A classic question to ask in such a logical study
would be "what formula characterizes this interrelationship between these ac­
cessibility relations."

It goes without saying that CORA will not be able to capture every nuance of
the mental states under discussion. Belief, desire, and intention are in reality
far too subtle, intricate, and fuzzy to be completely captured in a logic like
CORA. If such a theory was our goal, then the formalism would fail to satisfy
it. However, the logic is emphaticallynot intended to serve as such a theory.
Indeed, it seems certain that any formal theory which did fully capture all
nuances of belief, desire, and intention in humans would be of curiosity value
only: it would in all likelihood be too complex and involved to be of much use
for anything, let alone for actually building artificial agents. Instead of serving
as a "complete" theory of rational agency, therefore, the analysis in this chapter
is intended to investigate the properties that anidealized rational agent might
exhibit.
It is also worth noting that we are unable to answer many important
questions that deserve to be answered about the abstract properties of our logic.
In particular, complete axiomatizations - a rite of passage for any logic -
are not considered. The state of the art in multimodal logics is not sufficiently
92 Chapter 5

advanced that we can give a complete axiomatization of CORA. As pointed

out in the introduction to this book, this is not our goal. The intention instead is
to show how CORA can bea p plied to the problem of reasoning about rational
agents.
The chapter is structured as follows. The following section establishes a
number ofBDI co r res ponden ce results. These results give us a model-theoretic
characterization of various general interaction schemas. Using these results, I
go on to show how we can characterize possible pai r wise inte ra ctions between
the components of an agent's mental states.
Throughout this chapter, I USe the term "A-formula" to refer to formulae of
CORA that contain no positive occurrenCeS of the "E" path quantifier outside
the scope of a belief, desire, or intention modality. Clearly, A-formulae are
those that expressinevitabilities. Without loss of generality, I assume that all
A-formulae are written in the standard form A('P). Similarly, I USe the term
"E-formula" to refer to formulae that contain no positive occurrences of the
"A" path quantifier outside the scope of a belief, desire, or intention modality.
It should be obvious that E-formulae expresso ptional properties - properties
that are true in some (but not necessarily all) futures. As with A-formulae,
I assume that all E-formulae are written in the standard form E('P). Note
that Rao and Georgeff use "I-formula" and "O-formula" (for "inevitable" and
"optional") in the same way that I use "A-formula" and "E-formula" [185].

5.1 BDI Correspondence Theory*

Correspondence theory is perhaps the main reason for the success of Kripke­
style possible worlds semantics for modal logic [31]. Correspondence theory
investigates the relationships between properties of the accessibility relations
in Kripke models for modal logic, and axioms of the corresponding logic.
Most results take the form "formula schema 'P is valid in a model M iff the
accessibility relation R in M has property P." For example, in normal modal
logic, it is well-known that the formula schema

0 'P => 'P

is valid in a model iff the accessibility relation R that is used to give a semantics
to the "0" connective is reflexive (see, e.g., [31, p.80]).We can draw upon this
work to establish some basic properties of CORA - indeed, we already did
so, in the preceding chapter, to establish properties of the belief, desire, and
Properties of Rational Agents 93

intention modalities.
For CORA, however, we can extend the usual correspondence results
in several ways. Most obviously, we have available to us not just one, but
three modalities - Bel , Des, and Int . Multimodal logics were not widely
studied in classical correspondence theory, although given their importance in
contemporary computer science, they are studied increasingly often [29, 139].
Also, classical correspondence results assume that the worlds over which
accessibility relations operate have no internal structure. For CORA, however,
this is not the case: worlds are themselves branching time structures. Because
of this, we can investigate st ru ctu ral correspondence results, which in addi­
tion to incorporating interrelationships between the accessibility relations that
characterize the modal connectives, also incorporate structural properties of
worlds.

Subworlds

The most obvious structural relationship that can exist between two worlds
- and the most important for our purposes - is that of one world being a
sub wo rld of another. Intuitively, a world w is said to be a subworld of world
w' if w has the same structure as w' but has fewer paths. Formally, if w,w'
are worlds, then w is a subworld of w' iff paths ( w) � paths ( w'). If w is a
subworld ofw', then we indicate this by writingw�w'. The subworld relation
is illustrated in Figure 5.l.
It is not hard to see that ifw�w', then any A-formula that is true inw' will
also be true inw. Similarly, any E-formula that is true inw must also be true in
w'. These two facts tum out to be very useful in the analysis that follows, and
so we prove them in the following lemmas.

LEMMA 5.1: Suppose (M,V,w,t) is an interpretation, E(cp) is an E­

formula, and w,w' are two worlds in M such that t E Tw and t E TW'. Then if
(M,V,w,t) FS E(cp) and w�w' then (M,V,w',t) FS E(cp).

P roof Assume(M,V,w,t) FS E(cp).We need to show that ifw�w',t E Tw

and t E TW1 then (M,V,w',t) FS E(cp). From the semantics of E, we know
that (M,V,w,p) F'P cp for some path p E paths ( w) such that p(O) t. But if
=

w � w' then p E paths ( w'). Hence (M,V,w',p) F'P cp, and since p(O) t, =

by the semantics of E we have (M,V,w',t) FS E(cp) and we are done. _

94 Chapter 5

vv

vv'

vv"

("no",,")

Figure 5.1
' ' " '
The suhworld relationship: w is a suhworld of world w, that is, w � w; similarly, w � w.

LEMMA 5.2: Suppose (M, V, w,t) is an interpretation, A cp

( ) is an A­
fonnula, and w, w' are two worlds in M such thatt E Tw andt E TW'. Then if
(M, V, w,t) FS A cp
( ) and w' � w then (M, V, w',t) FS A cp
( ).

P roof Assume (M, V, w,t) FS A cp ( ) and thatt E Tw andt E TW'. We need

to show that if w' � w then (M, V, w',t) FS A cp( ). From the semantics of
A , we know that (M, V, w, p) FP cp for all paths p E paths ( w) such that
p (O) =t. But if w' � w then if p E paths ( w) then p E paths ( w'). Hence
Properties of Rational Agents 95

(M,V, w', p) FP cp, for all p E paths ( w/) such that p (O) t, and so from the
=

semantics of A , we have (M,V, w',t) FS A (cp) and we are done. _

Subset Correspondence Results

Given two accessibility relations R and R, the most obvious relationships

that could exist between them are subset relationshi ps. There are three main
possibilities.

R ([i]) � R ([i]) (5.1)

R ([i]) � R ([i]) (5.2)
R ([i]) = R ([i]) (5.3)

We can easily obtain a characterization of these relations in terms of the

.cORA formulae that they correspond to.

THEOREM 5.1: Suppose M and IVI are two of the modal connectives Bel ,
Des , and Int such that M ¥- lVI, defined by accessibility relations R and R
respectively. Then if we have some model in which R ([i]) � R ([i]), then
(Mi cp) :::} (IVI i cp) is valid in that model.

P roof Assume R ([i]) � R ([i]), and (M,V, w,t) FS (M i cp) for arbitrary
(M,V, w,t). We need to show that (M,V, w,t) FS (IVI i cp). From the seman­
tics of M, we know that (M,V, w',t) FS cp for all w' E R7 ([i]). Now since
R7 ([i]) � R7 ([i]), then (M,V, w",t) FS cp for all w" E W such that w" E
R7 ([i]). Hence from the semantics for lVI, we have (M,V, w,t) FS (IVI i cp)
and we are done. _

Given this theorem, it is simple to establish the following.

THEOREM 5.2: Suppose M and IVI are two of the modal connectives Bel ,
Des , and Int such that M ¥- lVI, defined by accessibility relations R and R
respectively. Then if we have some model in which R ([i]) =R ([i]), then
) )
(Mi cp ¢:} (IVI i cp is valid in that model.

P roof Observe that R ([i]) R ([i]) iff both R ([i]) � R ([i]) and R ([i])
= C
R ([i]). The result follows immediately from Theorem 5.1. _
96 Chapter 5

The next properties we investigate are obtained by combining the subset

relationships (5.1)-(5.3) with the subworld relationship that was introduced
above.
The first property we consider is thest ru ctu ral subset relationship between
accessibility relations.We say that accessibility relation R is a structural subset
of accessibility relation R if for every R-accessible world w, there is an R­
accessible world w' such that w is a subworld of w'. Formally, if R and R
are two accessibility relations then we write R �sub R to indicate that if
'
w E R7([i]), then there exists some w" E R7([i]) such that w' � w". If
R �sub R, then we say R is a st ru ctu ral subset of R. The structural subset
relationship is illustrated in Figure 5.2. In this case, we have three worlds, w,
' " " ' '
w , and w , such that w � w � w. Moreover, we have w E B� ([i]) and
"
w E 'D� ([i]). In this case, since for every desire-accessible world w' there
is a belief-accessible world w" such that w' � w", the desire accessibility
relation 'D is a structural subset of the belief accessibility relation B.
The following theorem characterizes the properties of structural subsets.

THEOREM 5.3: Suppose M andM are two of the modal connectives Bel ,
Des, and Int such that M "# M, defined by accessibility relations R and R
respectively, such that R7([i]) �sub R7([i]). Then for any A-formula A <p
( ),
we have M( i A <p
( )) => (M i A <p
( )) is valid in that model.

P roof If R7([i]) �sub R7([i]), then for every w' E R7([i]), there is some
"
w E R7([i]) such that w' � w". Now assume (M,V, w,t) FS M ( ( ))
i A <p
( )) for some (M,V, w,t). Then there must be
and (M,V, w,t) FS ...., (M i A <p
some w' E R7([i]) such that(M,V, w',t) FS ...., A <p
( ). But since R �sub R, then
" ' "
there is some w E R7([i]) such that w � w . Now since (M,V, w,t) FS
"
M
( ( )) then from the semantics ofM, it follows that (M,V, w ,t) FS
i A <p
'
( ) and thus by Lemma 5.2, (M,V, w ,t) FS A <p
A <p ( ). This is a contradiction,
so the assumption must be false. _

We write R �sup R to indicate that if w' E R7([i]), then there exists

some w" E R7([i]) such that w" � w'. If R �sup R, then we say R is a
st ru ctu ral su pe rset of R. In other words, if R is a structural superset of R, then
for every R-accessible world w, there is an R-accessible world w' such that w'
is a subworld of w. In Figure 5.2, belief is a structural superset of desire. As
with Theorem 5.3, we can establish a result that characterizes the structural
superset property.
Properties of Rational Agents 97

w'

, ?
----------- -
-
-
,
,

,
,
,
,

" -
- p

Figure 5.2
Illustrating structural subset/superset relationships: desire is a structural subset of belief
(D �sllb 8) and conversely, belief is a structural superset of desire (8 �SIIp D).

THEOREM 5.4: Suppose M and IVI are two of the modal connectives Bel ,
Des , and Int such that M ¥- lVI, defined by accessibility relations Rand R
respectively, such that R c;;,sup R. Then for any E-formula E('P) , we have
(IVI i E('P) ) => (M i E('P) ) is valid in that model.

P roof Similar to Theorem 5.3. •

Intersection-style Correspondences

Consider the following formula schema of CORA.

(Bel i 'P) => .(Bel i.'P)

98 Chapter 5

This formula schema is an instance of the "D" axiom from normal modal
logic, which corresponds to the accessibility relation being serial [31, p.80].
It is usually understood as expressing the consistency of belief. If we tum our
attention to the multimodal setting of .cORA, then we can consider related
schemas, which capture consistencybet ween modalities. These axiom schemas
have the following general form.

It turns out that these schemas are captured by theinte rse ction of the respective
accessibility relations being non-empty.

THEOREM 5.5: Suppose M and rill are two of the modal connectives Bel ,
Des, and Int such that M "I- rill, defined by accessibility relations R and R
respectively. Then if we have some model in which R7 ([i]) n R7 ([i]) "I- 0,
then (M i 'P) ::::} -, (rill i -''P) is valid in that model, as is the contrapositive
(rilli 'P) ::::} -,(M h'P) .

P roof I only prove that (M i 'P) ::::} -, (rill i 'P) is valid; once we have
this, showing that (rill i 'P) ::::} -, (M i 'P) is valid is trivial (it is simply the
contrapositive). Start by assuming R7 ([i]) n R7 ([i]) "I- 0, and (M, V, w,t) FS
(Mi 'P) for arbitrary (M, V, w,t). Then for all worlds w' E R7 ([i]), we have
(M, V, w',t) FS 'P. We need to show that (M, V, w,t) FS -, ( riII i -''P) . So
assume the contrary, i.e., that (M, V, w,t) FS (rill i -''P) . Then for all worlds
w" E R7 ([i]), we have (M, V, w",t) FS -''P. But since R7 ([i]) n R7 ([i]) "I- 0,
then for some Will E W, such that Will E R7 ([i]) and Will E R7 ([i]), we
must have both (M, V, Will,t) FS 'P and (M, V, Will,t) FS -''P. This is a
contradiction, so the assumption must be false. _

Just as we were able to undertake a more fine-grained analysis of the basic

interactions among beliefs, desires, and intentions by considering the structure
of worlds, so we are also able to undertake a more fine-grained characterization
of multimodal consistency properties by taking into account the structure of
worlds.
First, we write R7 ([i]) nsup R7 ([i]) to denote the set of worlds w' E
R7 ([i]) for which there exists some world w" E R7 ([i]) such that w' �
w". Referring to Figure 5.2, we have D� ([i]) nsup B� ([i]) = { w"}, but
Properties of Rational Agents 99

B� ([i]) nsup 1J� ([i]) = 0. Formally, R7([i]) nsup R7([i]) is defined as follows.

R7([i]) nsup R7([i]) = {w' I w' E R7([i]) and3w" E R7([i]) s.t. w'� w"}

The following theorem captures the key properties ofnsup .

THEOREM 5.6: Suppose M and IVI are two of the modal connectives Bel ,
Des , and Int such that M "I- lVI, defined by accessibility relations R and R
respectively. Then if we have some model in which R7([i]) nsup R7([i]) "I- 0,
then (M i A( rp)) => ,(IVI hA(rp)) is valid in that model.

P roof Suppose not. Then for some(M,V, w,t) we have R7([i])nsupR7([i]) "I-
0, and (M,V, w,t) FS (M i A(rp)) and (M,V, w,t) FS (IVI i ,A(rp)). Now
since R7([i]) nsup R7([i]) "I- 0, there exists some w' E R7([i]) such that
there is some w" E R7([i]) where w' � w". Hence from the semantics of
M, we know (M,V, w",t) FS A(rp) and from the semantics of lVI, we know
(M,V, w',t) FS ,A(rp). But since w' � w", then from Lemma 5.2 we know
that (M,V, w',t) FS A(rp). This is a contradiction, so the assumption must be
false. _

We can define nsub in the obvious way.

R7 ([i]) nsub R7 ([i]) = {w' I w' E R7 ([i]) and3w" E R7([i]) s.t. w"� w'}

The result corresponding to Theorem 5.6 is easily obtained.

THEOREM 5.7: Suppose M and IVI are two of the modal connectives Bel ,
Des , and Int such that M "I- lVI, defined by accessibility relations R and R
respectively. Then if we have some model in which R7([i]) nsub R7([i]) "I- 0,
then (M i E(rp)) => ,(IVI i ,E(rp)) is valid in that model.

P roof Analogous to Theorem 5.6. _

5.2 Pairwise Interactions between Beliefs, Desires, and Intentions

Now that we have established some general correspondence results for

.cORA, we can start to exploit them by considering the possible pai r wise in­
te ra ctions between the modal connectives Bel , Des, and Int . Given two modal
connectives M and lVI, the most important interactions of interest between them
100 Chapter 5

Table 5.1
Pairwise subset interrelationships between BD! modalities.

Relationship Formula Schema

W([i]) � 77([i]) (Int icp) => (Des icp)
77([i]) � W([i]) (Des icp) => (Int icp)
W([i]) � 8mi]) (Bel icp) => (Des icp)
87([i]) � W([i]) (Des icp) => (Bel icp)
87([i]) � 77([i]) (Int icp) => (Bel icp)
77([i]) C 87([i]) (Bel icp) => (Int icp)

have the following structure.

=?
Mi
( cp) Mi
( cp) (5.4)
{:::

Theorem 5.1 immediately gives us a way to characterize interaction schemas

of the form (5.4). Given the three modal operators Bel , Des , and Int , there are
exactly six subset relationships that can exist among them, and six correspond­
ing formula schemas. These interrelationships are summarized in Table 5.1.
Let us consider each of these formula schemas in turn.

(Inticp) =? (Desicp) (5.5)

Schema (5.5) says that if an agent intends something, then it desires it. Intu­
itively, this schema makes sense for rational agents. If we return to the view
suggested in chapter 2, of desires as options, then (5.5) can be seen as a con­
straint on the agent's deliberation process - an agent should only commit to
something as an intention if it considered it an option.

(Desicp) =? (Inticp) (5.6)

Schema (5.6) says that if an agent desires something, then it intends it. In
other words, an agent intends all its options. If we accept (5.5) as capturing
a property of rational agents, then if we also adopted (5.6), then desire and
intention would be equivalent. Formula (5.6) does not therefore appear to
capture any interesting properties of agents.

(Bel icp) =? (Desicp) (5.7)

Schema (5.7) is a well-known, if not widely-admired property of agents known

as realism. It was introduced by Cohen and Levesque [35], where, it was
argued, it captured the idea of an agent "accepting the inevitable." For example,
Properties of Rational Agents 10 1

suppose I believe that the sun will definitely rise tomorrow. Then, one could
argue, it makes no sense for me to desire that the sun will not rise. For whatever
I do, the sun will indeed rise tomorrow. So I may as well accept this as a desire.
As a property of rational agents, realism seems too strong. Later in this chapter,
we see how a finer grained analysis of the interrelationships among beliefs,
desires, and intentions can lead to more reasonable types of realism.

(Desi'P) => (Bel i'P) (5.8)

Schema (5.8) says that if an agent desires something, then it believes it. To give
a concrete example, suppose I desire I am rich: should I then believe I am rich?
Clearly not.

(Inti'P) => (Bel i'P) (5.9)

Schema (5.9) says that if an agent intends something, then it believes it. Let us
make this example concrete. Suppose I have an intention to write a book; does
this imply I believe I will write it? One could argue that, in general, it is too
strong a requirement for a rational agent. For example, while I can envisage
many possible futures in which I proudly present my finished manuscript to a
grateful and generous publisher, I can envisage equally many others in which
this happy event does not take place. (The fact that you are reading this, of
course, suggests that I was successful in my intention.) For example, I might
win the lottery next week, in which case I would certainly reconsider my
intention to write a book. Alternatively, I might be run down by a speeding
bus, and so on. So while I certainly believe it is possible that I will succeed
in my intention to write the book, I do not believe it is inevitable that I will
do so. Bratman [20, p.31] argues that if an agent has an intention to do some
action, then it is rational of the agent to believe that it will achieve it, given
certain background conditions - he refers to this property asintention-belief
consistency.

(Bel i'P) => (Inti'P) (5.10)

Schema (5.10) says that if an agent believes something, then it intends it. Like
(5.7), this is a kind of realism property, although it should be immediately ap­
parent that it expresses an even stronger, and arguably less desirable property.
To return to the example above, suppose that I believe that'P is true: should
I then adopt'P as an intention? Clearly not. This would imply that I would
choose and commit to everything that I believed was true. Intending some-
102 Chapter 5

Table 5.2
Pairwise structural subset interrelationships.

Relationship Formula Schema

'D([i]) c;.slIb I([i]) (Int iA(<p)) => (Des iA(<p))
I([i]) c;.SI,b 'D([i]) (Des iA(<p)) => (Int iA(<p))
'D([i]) c;.slIb B([i]) (Bel iA(<p)) => (Des iA(<p))
B([i]) c;.Sllb 'D([i]) (Des iA(<p)) => (Bel iA(<p))
B([i]) c;.SI,b I([i]) (Int iA(<p)) => (Bel i A(<p))
I([i]) C SI,b B([i]) (Bel iA(<p)) => (Int i A(<p))

thing implies selecting it and committing resources to achieving it. It makes no

sense to suggest committing resources to achieving something that is already
true.
To summarize, only formula (5.5) seems to capture a reasonable property
of rational agents. While in their current form they are too strong, we can also
see that weaker versions of (5.8) and (5.9) arguably also capture properties of
rational agents.
Formulas (5.5)-(5.10) are a useful starting point for our analysis of the pos­
sible relationships that exist among the three components of an agent's mental
state. However, it is clear that a fine r-grained analysis of the relationships is
likely to yield more intuitively reasonable results.

Attitudes to Inevitabilities

In our framework, we can also consider two important variations of (5.4).

The first relates an agent's attitudes to A-formulae, i.e., formula that express
inevitable futures.
=?
(Mi A 'P))
( M
( i A 'P))
( (5.11)
{:::

The six possible interaction schemas implied by Theorem 5.3 are summarized
in Table 5.2. Consider first (5.12).

(Inti A 'P
( )) =? (Desi A 'P))
( (5.12)

This says that if an agent intends that'P is inevitably true, then it desires that'P
is inevitably true. This schema, a variant of (5.5), is obviously an appropriate
property to demand of rational agents.

(Desi A 'P
( )) =? (Inti A 'P))
( (5.13)
Properties of Rational Agents 103

Schema (5.13) says that if an agent desires thatcp is inevitably true, then it
intends thatcp is inevitably true. In other words,cp is inevitable in all the options
considered by the agent. For example, suppose that in all the options that I
consider, it is inevitable that I am married. Then arguably, I should adopt an
intention to become inevitably married. So schema (5.13) seems to capture a
reasonable property of rational agents.

(Bel i A cp
( )) :::} ( ))
(Des i A cp (5.14)

Schema (5.14) says that if an agent believescp is inevitably true, then it

desires it is inevitably true. This schema is a variant of the undesirable realism
schemata, (5.7). It can be rejected as a property of rational agents for the same
reason as (5.7).

(Des i A cp
( )) :::} ( ))
(Bel i A cp (5.15)

Schema (5.15) says that if an agent desires something is inevitably true, then
it believes it is inevitably true. To be concrete, suppose that in all the options
I consider, it is inevitable that I am married. Do I then believe I am inevitably
married? No. I may well believe there is some future in which I am not married,
even though in all the options I consider, it is inevitable that I am married.

(Int i A cp
( )) :::} ( ))
(Bel i A cp (5.16)

Schema (5.16) says that if an agent intends thatcp is inevitably true, then it
believes it is inevitably true. As we noted earlier, Bratman argues that intending
cp certainly provides some support for the belief thatcp [20, p.38]. Since
intentions are in some SenSe conduct-controlling, then all other things being
equal, my intention to achievecp must makecp more likely than it would be if I
did not intendcpo However, this is not to say that my intendingcp is inevitable
implies I believecp is inevitable. So while (5.16) is arguably more reasonable
than (5.9), it is still too strong to be a necessary property of rational agents.

(Bel i A cp
( )) :::} (Int i A cp
( )) (5.17)

Schema (5.17) says that if an agent believescp is inevitably true, then it intends
it is inevitably true. For the same reasons as (5.10), we reject this schema as a
property of rational agents.
104 Chapter 5

Table 5.3
Pairwise structural superset interrelationships between modalities.

Relationship Formula Schema

'D([i]) r;;.sup I([i]) (Int iE(cp)) => (Des iE(cp))
I([i]) r;;.SlIP 'D([i]) (Des iE(cp)) => (Int iE(cp))
'D([i]) r;;.sup B([i]) (Bel iE(cp)) => (Des iE(cp))
B([i]) r;;.SlIP 'D([i]) (Des iE(cp)) => (Bel iE(cp))
B([i]) r;;.SlIP I([i]) (Int iE(cp)) => (Bel iE(cp))
I([i]) C Sllp B([i]) (Bel iE(cp)) => (Int iE(cp))

Attitudes to Options

In the preceding section, we considered the possible attitudes an agent could

have to inevitable states of affairs. Next, we consider an agent's attitudes to
o ptions.

=?
(Mi E cp
( )) M
( i E cp
( )) (5.18)
{:::

As with Theorem 5.1, Theorem 5.4 immediately gives us six possible interac­
tions, summarized in Table 5.3.
Consider first (5.19).

(Inti E cp
( )) =? (Desi E cp
( )) (5.19)

In order to understand this schema, it is worth considering what it means to

have an intention towards an E-formula such as E cp ( ). Intuitively, an agent with
such an intention wants to "keep its options open." In other words, it intends to
ensure thatcp is always true on at least one possible future - in other words,
thatcp is never ruled out. There are many obvious situations where such an
intention is appropriate. For example, although I have a university job, I might
want to keep the option open of going into industry and taking a proper job,
while not having the intention to make a job in industryinevitable. Intentions
towards E-formulae capture exactly such commitments.
Returning to (5.19), this schema says that if an agent intends thatcp is
optionally true, then it desires thatcp is optionally true. Clearly such an schema
makes sense for rational agents.

(Desi E cp
( )) =? (Inti E cp
( )) (5.20)
Properties of Rational Agents 105

Schema (5.20) says that if an agent desires that<p is true on at least one
future, then it intends that<p is true on at least one future. This formula can
be understood as characterizing an agent who wants to commit to all options;
in other words, an agent who will never rule out options. Such a property
might make sense in some situations. For example, suppose I have decided
what my options are, but have insufficient information to make an informed
choice among my options.

( ))
(Bel i E <p :::} ( ))
(Des i E <p (5.21)

Schema (5.21) says that if an agent believes<p is true on at least one future, then
it desires it is true on at least one future. This schema can be seen as character­
izing a weaker variant of the realism property (5.7) from the preceding section.
However, it still seems in some sense too strong. For example, suppose I be­
lieve there is some future in which there will be a nuclear war. Do I also desire
that there is some future in which there is a nuclear war? Clearly not. So, like
the realism schemas (5.7) and (5.10), we reject (5.21) as a desirable property
of rational agents.

( ))
(Des i E <p :::} ( ))
(Bel i E <p (5.22)

Schema (5.22) says that if an agent desires something is optionally true, then
it believes it is optionally true. This formula is a variant of (5.8). However,
unlike (5.8), schema (5.22) seems an entirely reasonable property for rational
agents. For if I desired that I was rich in some future, while not believing that
there was any possible future in which I was rich, then I would clearly be
irrational.

( ))
(Int i E <p :::} ( ))
(Bel i E <p (5.23)

Schema (5.23) says that if an agent intends that<p is optionally true, then it
believes it is optionally true. For similar reasons to (5.22), this schema is a
reasonable property of rational agents.

( ))
(Bel i E <p :::} ( ))
(Int i E <p (5.24)

Schema (5.24) says that if an agent believes<p is optionally true, then it intends
it is optionally true. As with (5.21), this is not a property we would demand of
rational agents.
106 Chapter 5

Table 5.4
Weak realism interactions.
Relationship Formula Schemas
W([i]) n1:;"([i]) i-0 (Int icp) => ,(Des i ,cp)
1:;"([i]) n W([i]) i-0 (Des icp) => ,(Int i ,cp)
B;"([i]) n W([i]) i-0 (Bel icp) => ,(Des i ,cp)
W([i]) nB7([i]) i-0 (Des icp) => ,(Bel i ,cp)
1:;"([i]) nB7([i]) i-0 (Int icp) => ,(Bel i ,cp)
B;"([i]) n1:;"([i]) i-0 (Bel icp) => ,(Int i ,cp)

5.3 Varieties of Realism

We saw above that realism properties such as (Bel i r.p) => (Des i r.p) are in
general too strong to be regarded as reasonable properties of rational agents.
In this section we will consider some weaker realism properties.
As before, this general result enables a number of specific interaction
properties to be established - see Table 5.4.
The first of these properties, represented by schemas (5.25) and (5.26),
says that an agent's intentions are consistent with its desires, and conversely,
its desires are consistent with its intentions.

(Inti r.p) => ,(Desi 'r.p) (5.25)

(Desi r.p) => ,(Inti 'r.p) (5.26)

These schemas, which captureintention-desi re consistency, appear to be rea­

sonable properties to demand of rational agents in some, but not all circum­
stances. Intention-desire consistency states that an agent only ever considers
options that are consistent with its intentions. Under normal circumstances, we
might consider such behavior acceptable, or even desirable. For example, if I
have an intention to r.p then it will not normally make sense for me to entertain
options that are inconsistent with r.p. However, under certain circumstances,
it makes sense for an agent to re conside r its intentions - to deliberate over
them, and possibly change focus. This implies entertaining options that are not
necessarily consistent with current intentions. An agent that always respects
schemas (5.25) and (5.26) will never do such reconsideration.
Schemas (5.27) and (5.28) capturebelief-desi re consistency.

(Bel i r.p) => ,(Desi 'r.p) (5.27)

(Desi r.p) => ,(Bel i 'r.p) (5.28)

Properties of Rational Agents 107

The first of the belief-desire consistency schemas, (5.27), is also known as

belief-desi re weak realism. Recall that belief-desire realism, as represented by
schema (5.7) was too strong to be an acceptable version of realism. Schema
(5.27) captures a weaker notion of belief-desire realism. It simply says that if
an agent believesrp is already true, then an agent should not desire thatrp is not
true. Thus if I believe it is raining, there is no point in desiring it is not raining,
since I will not be able to change what is already the case.
The second belief-desire consistency schema, (5.28), says that if an agent
desiresrp, then it does not believerp is false. On first consideration, this schema
seems unreasonable. For example, I may desire to be rich while believing that
I am not currently rich. But when we distinguish between p resent-di re cted
and futu re-di re cted desires and beliefs, the property makes sense for rational
agents. It is not rational for me to have a desire to be rich at this moment
while simultaneously believing that I am not rich at this moment. However, it is
entirely acceptable for me to simultaneously believe that I am not rich currently
while desiring to be rich at some time in the future. The second belief-desire
consistency schema, (5.28), rules out the first case, but has nothing to say about
the second.

(Intirp) => ...,(Bel i...,rp) (5.29)

(Bel irp) => ...,(Inti...,rp) (5.30)

Similar comments apply to the intention-belief consistency schemas, (5.29)

and (5.30), as to the belief-desire consistency schemas, and we will therefore
not comment on them here.

Weak Realism and Inevitabilities

Strong realism, as discussed above, is the most general kind of realism that
we might consider. However, for the reasons we discussed above, most strong
types of realism are not acceptable properties of rational agents. In this and the
subsection that follows, we will discuss realism with respect toinevitable and
o ptional formulae.
Theorem 5.6 generates the results shown in Table 5.5.
The first of these properties, (5.31), says that if an agent intends to bring
about A rp ( ); schema (5.32) expresses the
( ), then it should not desire that..., A rp
same property for desires and intentions.

(Inti A rp
( )) => ...,(Desi..., A rp
( )) (5.31)
108 Chapter 5

Table 5.5
Inevitabilities and weak realism properties.

Relationship Formula Schemas

:r;"W]) nSllp 1J7W]) i- 0 (Int iA(rp)) => ,(Des i ,A(rp))
W([i]) nSllp :r;"([i]) i-0 (Des iA(rp)) => ,(Int i ,A(rp))
13;"([i]) nSllp W([i])i-0 (Bel iA(rp)) => ,(Des i ,A(rp))
W([i]) nSllp 13;"([i])i-0 (Des iA(rp)) => ,(Bel i ,A(rp))
:r;"([i]) n.",p 13;"([i]) i-0 (Int iA(rp)) => ,(Bel i ,A(rp))
13;"([i]) nSllp :r;"([i]) i-0 (Bel iA(rp)) => ,(Int i ,A(rp))

(Des i A 'P))
( :::} ,(Int i , A 'P
( )) (5.32)

Schema (5.33) says that if an agent believes'P is inevitable, then it should not
desire it is not inevitable. This captures an acceptable realism property.

(Bel i A 'P))
( :::} ,(Des i , A 'P
( )) (5.33)

Schema (5.34), states that if an agent desires something is inevitable, it should

not believe it is not inevitable; this Seems an acceptable, though not necessarily
desirable property of rational agents.

(Des i A 'P))
( :::} ,(Bel i , A 'P
( )) (5.34)

Finally, schemas (5.35) and (5.36) capture weak realism between intentions
and beliefs.

(Int i A 'P))
( :::} ,(Bel i , A 'P
( )) (5.35)

(Bel i A 'P))
( :::} ,(Int i , A 'P
( )) (5.36)

Weak Realism and Options

Table 5.6 summarizes the formula schemas corresponding to Theorem 5.7.

The first of these properties, (5.25), says that if an agent intends that'P
should be optionally true, then it should not desire that'P is not optionally true.
Clearly such consistency is a desirable property of rational agents.

(
(Int i E 'P)) :::} ( ))
,(Des i , E 'P (5.37)
Properties of Rational Agents 109

Table 5.6
Options and weak realism properties.

Relationship Formula Schemas

77([i]) nSllp W([i]) i- 0 (Int iE(cp)) => ,(Des i ,E(cp))
1))"([i]) nSl'P77([i]) i- 0 (Des iE(cp)) => ,(Int hE(cp))
8mi]) nSllp W([i]) i- 0 (Bel iE(cp)) => ,(Des i ,E(cp))
1))"([i]) nsup 87([i]) i- 0 (Des iE(cp)) => ,(Bel i ,E(cp))
77([i]) nSllp 87([;]) i- 0 (Int iE(cp)) => ,(Bel i ,E(cp))
87([;]) nSllp 77([i]) i- 0 (Bel iE(cp)) => ,(Int i ,E(cp))

Schema (5.38) says that if an agent desires thatcp is optional, then it does
not intend thatcp is not optional.

(Desi E cp
( )) ::::} -,(Inti -,E cp
( )) (5.38)

Schema (5.39) says that if an agent believescp is optional, it does not desire
thatcp is not optional.

(Beli E cp
( )) ::::} -,(Desi -,E cp
( )) (5.39)

Schema (5.39) says that if an agent desires thatcp is optional, then it does not
believecp is optional. Earlier, we argued that schema (5.7), which said that if
an agent believes something it also desires it, was unacceptably strong for most
agents. Schema (5.39), which can be seen as a weak variant of (5.7), is clearly
more acceptable, as is (5.40).

(Desi E cp
( )) ::::} -,(Bel i -,E cp
( )) (5.40)

Schemas (5.41) and (5.42) say that an agent's optional intentions are consistent
with its beliefs, and as with the desire case, above, these schemas are clearly
more acceptable than either (5.9) or (5.10).

(Inti E cp
( )) ::::} -,(Bel i -,E cp
( )) (5.41)

(Bel i E cp
( )) ::::} -,(Inti -,E cp
( )) (5.42)

5.4 System s of BDI Logic

So far, we have only been concerned with binary relationships between modal­
ities. But we can also consider ternary (i.e., three-way) relationships between
the modalities. Given the number of different binary relationships we have
1 10 Chapter 5

Table 5.7
Systems OfBOI logic. (Source: [185, p.321].)

Name Semantic Corresponding Formula Schema

Condition
BOI-SI B c;.sup V c;.sup I (Int iE(cp)) => (Des iE(cp)) => (Bel iE(cp))
BOI-S2 B c;.sub V c;.sub I (Int iA(cp)) => (Des iA(cp)) => (Bel i A(cp))
BOI-S3 B CV C I (Int icp) => (Des icp) => (Bel icp)
BOI-RI I c;.sup V c;.sup B (Bel iE(cp)) => (Des iE(cp)) => (Int iE(cp))
BOI-R2 I c;.Sllb V c;.Sllb B (Int iA(cp)) => (Des iA(cp)) => (Bel i A(cp))
BOI-R3 I C V CB (Bel icp) => (Des icp) => (Int icp)
BOI-WI B nsupVi-0 (Bel iA(cp)) => ,(Des i ,A(cp))
v nSl1pIi-0 (Des iA(cp)) => ,(Int i ,A(cp))
B nSl1pIi-0 (Bel iA(cp)) => ,(Int i ,A(cp))
BOI-W2 B nSllb Vi-0 (Bel iE(cp)) => ,(Des i ,E(cp))
V nSl1bIi-0 (Des iE(cp)) => ,(Int i ,E(cp))
B nSllb Ii-0 (Bel iE(cp)) => ,(Int i ,E(cp))
BOI-W3 B nvi-0 (Bel icp) => ,(Des i ,cp)
v nIi-0 (Des icp) => ,(Int i ,cp)
B nIi-0 (Bel icp) => ,(Int i ,cp)

looked at so far, there are clearly a great many ternary relationships that we
might possibly consider. However, most of these relationships are meaning­
less (or at best capture obscure and frankly uninteresting situations). The most
important relationships for our purposes are those summarized in Table 5.7.

5.5 The Side-Effect Problem

One of the most well-known problems in the theory of intention is that ofside
effe cts. The side-effect problem is illustrated by the following scenario.

Janine intends to visit the dentist in order to have a tooth pulled. She is aware that as a
consequence of having a tooth pulled, she will suffer pain. Does Janine intend to suffer
pain?

In this example, suffering pain is a side effect of having a tooth pulled; having
a tooth pulled and suffering pain come as a "package deal." However, it is
generally agreed (and in particular, Bratman convincingly argues [20, p.142]),
that rational agentsdo not have to intend the consequences of their intentions.
In other words, Janine can intend to have a tooth pulled, believing that this will
cause pain, without intending to suffer pain.
Properties of Rational Agents 111

In CORA, the side-effect problem can be captured by simply requiring

the following schema to be satisfiable.

(Inticp) 1\ (Bel icp =>'lj;) 1\ ....,(Inti'lj;) (5.43)

In other words, it should be possible for an agent to intendcp , believing that

it is inevitable thatcp =>'lj;, without intending that'lj;. The following are also
examples of side-effect schemas.

(Inticp) 1\ (Desicp =>'lj;) 1\ ....,(Inti'lj;) (5.44)

(Desicp) 1\ (Bel icp =>'lj;) 1\ ....,(Desi'lj;) (5.45)

We can prove the following general result about such consequential closure
schemas.

THEOREM 5.8: Suppose M and !VI are two of the modal connectives Bel ,
Des , and Int such that M -I- !VI, defined by accessibility relations R and R
respectively. Then if we have some model in which R ([i]) � R ([i]), then the
schema

(Micp) 1\ (!VIicp =>'lj;) 1\ ...., (Mi'lj;)

will be unsatisfiable in this model.

P roof Assume (M,V, w,t) FS (Mi cp) 1\ (!VI i cp =>'lj;) for some arbitrary
(M,V, w,t). If R ([i]) � R ([i]), then by Theorem 5.1, the schema (!VI i X) =>
(Mi X) will be valid. Hence (M,V, w,t) FS (Mi cp =>'lj;). By the K axiom
for the M operator and propositional reasoning, (M,V, w,t) FS (Mi'lj;). This
contradicts the assumption. _

Thus the consequential closure schema (5.43) will be unsatisfiable in a model

ifl; � B7·

5.6 Notes and Further Reading

The kind of analysis undertaken in this chapter - of understanding the rela­

tionships among semantic constraints on Kripke relations and the modal logic
formulae these constraints make valid - is known as co r res ponden ce theory,
and has a long tradition in modal logic. The basic correspondence properties
were identified by Kripke himself [126], and indeed the elegance of correspon-
1 12 Chapter 5

dence theory accounts in a large part for the continued success of this style of
semantics. Basic correspondence results are described in [101] and [31].
The study of interactions among multimodallogics, of the kind we under­
take in this chapter, is a relatively recent development. Laurent Catach's Ph.D.
work, summarized in [29], was one of the first systematic, abstract studies to
receive a wide audience. Catach established a correspondence theorem analo­
gous to the general d,l,m,n correspondence result of normal modal logic [31,
p.88]. More recently, multimodal logics have been found to have a wide variety
of applications, and this recognition has led to a rapid increase in the number
of systematic studies of multimodal systems. Logical systems with multiple
modalities of the same time are discussed in [54]. Alessio Lomuscio gives a
comprehensive study of interaction schemas in multimodal settings, generaliz­
ing Catach's results [139, 140].
The specific properties discussed in this chapter were originally identified
in Rao and Georgeff's BDI papers. The paper that "introduced" BDI logics
is [187] - it showed how BDI logic could be used to capture many of our
intuitions about intentions, with particular reference to the work of Cohen and
Levesque [35]. Axiomatizations of the asymmetry thesis, and a justification
for the temporal branching structure of BDI logic were presented in [186].
The "definitive" paper on Rao and Georgeff's formalism is [185]. This paper
comprehensively studies possible axiomatizations ofBDI logic, and gives some
completeness results for propositionalBDI logic by means of semantic tableaux
systems.
6 Collective Mental States

So far in this book, we have been considering the properties of individual

agents. In this chapter, we change the emphasis from individual agents to mul­
tiagent systems. I begin by discussing how we might go about characterizing
the properties of groups of agents. Starting with mutual belief, desire, and in­
tention, I go on to show how joint commitment can be defined.

6.1 Mutual Beliefs, Desires, and Intentions

The intuition behind mutual belief is straightforward. Consider a situation

where you and I are in discussion, and I say to you "The door is open."
Assuming there are no abnormal circumstances (i.e., you were paying attention
to what I was saying, you did not suddenly lose your hearing, and so on), then
the fact that I uttered this sentence will be mutually believed by both of us. To
better understand what is meant by this, suppose 'P is the statement

"Michael uttered the sentence 'The door is open.' ''

Then you will believe 'P.

(Bel you 'P)

Moreover, I will also believe 'P.

(Bel me 'P)

In addition, however, since the act of me uttering the sentence was immediately
visible to both of us, and both of us were witness to the fact that both of us were
witness to it, it is only rational for me to believe that you believe 'P.

(Bel me (Bel you 'P))

The situation is of course symmetric, so you will also believe that I believe 'P.

(Bel you (Bel me 'P))

114 Chapter 6

It is not hard to see that, if we both assume that the other observed the event,
then one must carry on this nesting of beliefs.
(Bel me (Bel you (Bel mecp) ) )
(Bel you (Bel me (Bel youcp) ) )
(Bel me (Bel you (Bel me (Bel youcp) ) ) )
(Bel you (Bel me (Bel you (Bel mecp) ) ) )
(Bel me (Bel you (Bel me (Bel you (Bel mecp) ) ) ) )
(Bel you (Bel me (Bel you (Bel me (Bel youcp) ) ) ) )
... and so on.
We refer to this state of affairs - where I believe, you believe, I believe that
you believe, you believe that I believe, I believe that you believe that I believe,
and so on - as mutual belief It follows that, in principle at least, mutual belief
is infinitary in nature. If we try to write down a formula that captures mutual
belief between two or more agents, then initially at least, it appears that this
formula must be infinitely long. (In fact, we will see that this is not necessarily
the case, as we can characterize mutual mental states in terms offixed points;
nevertheless, mutual beliefs are infinitary in nature.)
It is important to distinguish mutual belief from the (much simpler) case,
where we have simply I believe and you believe:

(Bel mecp) 1\ (Bel youcp)

Mutual belief expresses a much stronger property, which implies this weaker
kind of collective mental state.
As the informal example given above illustrates, mutual mental states play
an important role in communication. But, more than this, they tum out to
be important in coordination problems. To illustrate what is meant by this,
consider the following scenario, from [54, p.176].
Two divisions of an army, each commanded by a general, are camped on two hilltops
overlooking a valley. In the valley awaits the enemy. It is clear that if both divisions
attack the enemy simultaneously they will win the battle, while if only one division at­
tacks, it will be defeated. As a result, neither general will attack unless he is absolutely
sure that the other will attack with him. In particular, a general will not attack if he
receives no messages. The commanding general of the first division wishes to coordi­
nate a simultaneous attack (at some time the next day). The generals can communicate
only by means of messengers. Normally, it takes a messenger one hour to get from one
encampment to the other. However, it is possible that he will get lost in the dark or,
worse yet, be captured by the enemy. Fortunately, on this particular night, everything
goes smoothly. How long will it take them to coordinate an attack?
Collective Mental States 115

Suppose a messenger sent by general A reaches General B with a message saying

"attack at dawn." Should General B attack? Although the message was in fact delivered,
General A has no way of knowing that it was delivered. A must therefore consider
it possible that B did not receive the message (in which case B would definitely not
attack). Hence A will not attack given his current state of knowledge. Knowing this,
and not willing to risk attacking alone, B cannot attack solely based on receiving A's
message. Of course, B can try to improve matters by sending the messenger back to A
with an acknowledgment. When A receives this acknowledgment, can he then attack? A
here is in a similar position to the one B was in when he received the original message.
This time B does not know that the acknowledgment was delivered.

Intuitively, the two generals are trying to bring about a state where it is
mutually believed (or known) between them that the message to attack was
delivered. Each successive round of communication, even if successful, only
adds one level to the depth of nested belief. No amount of communication
is sufficient to bring about the infinite nesting that mutual belief requires. As
it turns out, if communication delivery is not guaranteed, then mutual belief
can never arise in such a scenario. Ultimately, this is because, no matter how
many messages and acknowledgments are sent, at least one of the generals will
always be uncertain about whether or not the last message was received.
One might ask about whether irifinite nesting of mutual belief is required.
Could the two generals agree between themselves beforehand to attack after,
say, only two acknowledgments? Assuming that they could meet beforehand
to come to such an agreement, then this would be feasible. But the point is that
whoever sent the last acknowledgment would be uncertain as to whether this
was received, and would hence be attacking while unsure as to whether it was
a coordinated attack or a doomed solo effort.
More pragmatically, of course, it should be noted that - as with all the
formalizations of mental states that appear in this book - mutual belief and its
kin represent idealizations of true states of affairs. In all but the most unusual
of circumstances, humans are unlikely to make use of mental states that are
nested to a depth of more than a few levels.
In order to formally define mutual belief, we will make use of some
auxiliary definitions. First, we write (E-Bel g 'P) to represent the fact that
everyone in the group denoted by g believes 'P.

(E-Bel g 'P) � Vi· (i E g) => (Bel i 'P)

116 Chapter 6

If k E IN such that k > 0, we define (E-Bel k g <p)inductively, as follows.

(E-Bel k g <p) � { (E-Bel

(E-Bel g <p)
g (E_Bel k-1 g
if k = 1
<p)) otherwise.
The mutual belief of <pin group g is then defined as follows.
�
(M-Bel g<p) 1\ (E-Bel k g<p)
k>O
In other words, (M-Bel g <p)is the same as the following infinitary formula.

Vi· (i E g) =>
(Bel i <p)A
(Bel i (E-Bel g <p))A
(Bel i (E-Bel g (E-Bel g <p)))A
... and so on.

It turns out that there are several ways that we can characterize mutual belief.
We can adopt a model-theoretic approach, and characterize mutual belief
in terms of each agent's belief accessibility relation. Alternatively, we can
characterize mutual belief as afixed point.

The Semantics of Mutual Belief*

Mutual belief has a natural and quite intuitive semantic characterization, in

terms of each agent's belief accessibility relation. In order to understand this
characterization, let g � DAg be some group of agents, ( w, t) be a situation,
and w' E W be a world. Then we say w' is B(g)-reachable from ( w, t) in k
steps if we can reach world w' from ( w, t) by following the arcs of the belief
accessibility relations for agents in g for no more than k steps. Formally, w' is
B(g)-reachable from ( w, t) in k steps iff there is some sequence ( Wl' . . . ,Wk)
of worlds such that Wl = W, Wk = w', and for each ( wu, wu+ d we have
Wu+l E B� (i) for some agent i E g.
u

THEOREM 6.1: (M, V, w,t) FS (E-Bd g <p) iff (M, V, w',t) FS <p for
'
every world w such that w' is B(g)-reachable from w in k steps.

Proof The proof is by induction on k. The base case, where k = 1, requires

that if (M, V, w, t) FS (E-BeI1 g <p) then (M, V, w', t) FS <p for every w' E
B�' (i) , for every i E g. The result follows easily from the definition of E-BeI1.
For the inductive assumption, suppose that (M, V, w, t) FS (E-Belm g <p) iff
Collective Mental States 117

(M, V, w', t) I=s cp for every world w'such that w'is B (g) -reachable from w
in m steps. We need to show that in this case, (M, V, w, t) I=s (E_Bel mH gcp)
iff (M, V, w', t) I=s cp for every world w'such that w'is B (g) -reachable from
w in m + 1 steps. This follows trivially from the definition of E-Bel and the
inductive assumption. _

We will also say that w'is B (g) -reachable from (w, t) if w'is B (g) -reachable
from (w, t) in k steps, for some k E IN such that k > O. We can then prove the
following.

THEOREM 6.2: (M, V, w, t) I=s (M-Bel g cp) iff (M, V, w', t) I=s cp for
every world w'such that w'is B (g) -reachable from w.

Proof Follows immediately from Theorem 6.1 and the definition of M-Bel . _

Theorems 6.1 and 6.2 characterize E-Bel k and M-Bel graph theoretically,
where each agent's accessibility relation is interpreted as a graph over possible
worlds.

Mutual Mental States as Fixed Points*

The theory of fixed points is widely used in computer science, particularly in

the semantics of programming languages. A fixed-point equation is one with
the following form.

x = f(x)

Fixed-point equations like this often cause confusion, because they appear to
be in some sense circular - the left-hand side appears to be defined in terms
of the right. But such equations can have quite natural solutions. For example,
the equation x = x2 has solutions x = 0 and x = 1. The following Theorem
tells us that mutual belief can be understood as a fixed point.

THEOREM 6.3: I=s (M-Bel gcp) ¢:} (E-Bel gcp /\ (M-Bel gcp) ) .

Proof For the left to right implication, assume (M, V, w, t) I=s (M-Bel gcp) .
We need to show that this implies (M, V, w, t) I=s (E-Bel gcp/\ (M-Bel gcp) ) .
From Theorem 6.2, we know that (M, V, w', t) I=s cp for each w'that is B (g) ­
reachable from w. Now consider an arbitrary w" E B7(i) for some i E g.
Then w" is B (g) -reachable from w in one step, thus (M, V, w", t) I=s cpo It
118 Chapter 6

follows that (M, V, w, t) FS (Bel i <p) for all i E g, and so (M, V, w, t) FS

(E-Bel g <p). Now, consider any Will E W such that Will is B(g) reachable
from w". It must be that Will is also B(g)-reachable from w. It follows that
(M, V, Will, t) FS <p, and so (M, V, w", t) FS (M-Bel g <p). We conclude
that (M, V, w, t) FS (E-Bel g<pA (M-Bel g<p)). The right to left implication
follows from the inductive definition of M-Bel in terms of E-Bel . _

This Theorem tells us that (M-Bel g <p) can be understood as a fixed point of
the function

f (x ) = (E-Bel g<pA x)

where the functionf maps formulae of CORA to formulae of CORA. (See,

for example, [54, pp.402--411] for a detailed analysis of common knowledge
defined as a fixed point.)

Properties of Mutual Belief

In addition to enjoying the properties discussed above, mutual belief also

behaves like a KD4 modal operator.

THEOREM 6.4:

1. FS (M-Bel g<p ::::} �) ::::} ((M-Bel g<p) ::::} (M-Bel g�))

2. FS (M-Bel g<p) ::::} ...,(M-Bel g...,<p)
3. FS (M-Bel g<p)::::} (M-Bel g(M-Bel g<p))
4. IfFS <pthenFS (M-Bel g<p)

Proof Parts (1) and (4) are trivial; for (2) and (3), notice that the B(g)
reachability relation will be serial and transitive. (See e.g., [54, p.34].) _

In the remainder of the book, I will assume that M-Int (mutual intention) and
M-Des (mutual desire) have been defined in an analogous way to M-Bel .

6.2 Mutual Mental States and Teamwork

Mutual beliefs, desires, or intentions characterize states that exist in certain

idealized scenarios. In particular, as the informal example that opens this
chapter illustrates, mutual belief appears to characterize many communication
Collective Mental States 119

scenarios. In this section, we will consider the types of mutual mental states
that characterize real-world joint working scenarios, and in particular, the type
of mental states that characterize an agent engaged in teamwork. By teamwork,
we mean scenarios such as the following, where a group of agents work
together [133, 37, 250]:

• to lift a heavy object;

• to write a joint paper; or
• to drive in a convoy.

What sort of mental state characterizes such cooperative efforts? One possi­
bility would be to argue that there was a "team state" that was not reducible
to the individual beliefs, desires, and intentions of the participating agents. We
could easily formalize such a concept, by introducing (yet) another primitive
modal operator to represent team mental states, and then investigate its prop­
erties using the same techniques employed throughout this book. The problem
is that intuition rejects such an approach. It seems almost superstitious to ar­
gue that there is something about team action that exists independently of the
team members. I therefore adopt a reductionist approach, where teamwork is
analyzed in terms of the individual mental states of team members.
As a first attempt at defining team mental states, we might imagine that it
is sufficient for a team to have (say) a mutual intention (M-Int) towards some
particular goal. Suppose we have two agents i and 1 cooperatively working to
achieve some state of affairs <po Then (M-Int {i,l} <p)would expand to

(E-Int {i,l} <p)1\

(E-Int {i,l} (E-Int {i,l} <p))1\
(E-Int {i,l} (E-Int {i,l} (E-Int {i,l} <p)))1\
. . . and so on.
Thus agent i's mental state is characterized by the following pattern of inten­
tions.
(Int i <p)1\
(Int i (E-Int {i,l} <p))1\
(Int i (E-Int {i,l} (E-Int {i,l} <p)))1\
. . . and so on.
This pattern is symmetric, and so is shared by each participating agent. It seems
reasonable to suppose that an agent engaged in team activity does indeed have
such a pattern of intentions. For example, if you and I are together lifting a
120 Chapter 6

heavy object, it seems reasonable that I intend to lift it, and I intend that you
intend to lift it, and I intend that you intend that I intend to lift it, and so on.
But is there more to cooperative action than this? Levesque, Cohen, and Nunes
argue that such models of mutual intention fail to capture some important
characteristics of teamwork [133, 37]. In particular, they fail to capture the
notion of joint commitment. To continue with the example of lifting a heavy
object, suppose that while we are lifting it I come to believe that we will be
unable to lift it for some reason. As we saw in chapter 5, a rational agent acting
alone in such circumstances would drop the intention: it would be irrational
for an agent to act towards an intention that was believed to be impossible.
But in the teamwork case, should I simply drop my intention to lift the object
with you, because I believe it is not achievable? You would hardly be inclined
to say I was "cooperative" if I did so. What appears to be missing is that
associated with team action is what Jennings calls responsibility toward the
team [104, 105, 106]. Responsibility in the context of teamwork is comprised
of two parts: a commitment and a convention.
A commitment is a pledge or promise that an agent makes to the group in
which it is working. A commitment must, by definition, persist over time [35,
pp.217-219].
A convention is a means of monitoring a commitment. A convention thus
specifies the circumstances under which a commitment can be abandoned
and how an agent should behave both locally and towards others when one
of these conditions arises [105]. Thus a convention identifies the conditions
under which the joint commitment can be dropped, and also describes how the
agent should behave towards its fellow team members. For example, if an agent
drops its joint commitment because it believes that the intention will never be
attained, then it is part of the notion of "cooperativeness" that it informs fellow
team members of this. Conventions thus provide a commonly understood
frame of reference in which agents can work. By adopting a convention, every
agent knows what is expected both of it, and of every other agent, as part of
the collective working towards the intention, and knows that every other agent
has a similar set of expectations. By adopting a particular convention, team
behavior becomes more predictable. Predictability is one of the most important
enablers of coordinated behavior [18].
We can formalize conventions quite simply. A convention is a set of rules,
where each rule r = (p, 'Y) is a pair consisting of a re-evaluation condition p
and an associated goal 'Y. A convention will be adopted by a group of agents g
when working towards some goal 'P. The idea is that if ever an agent i E g
Collective Mental States 121

believes p to be true, then it must adopt 'Y as an intention, and keep this
intention until the commitment becomes redundant.
Formally, a convention, c, is a finite indexed set of pairs:

where Pk is a state formula of CORA called the re-evaluation condition, and

'Yk is a state formula of CORA called goal.
Team or joint commitments have a number of parameters. First, a team
commitment is held by a group g of agents. Second, joint commitments are
held with respect to some goal rp; this is the state of affairs that the group is
committed to bringing about. They also have a pre-condition, which describes
what must initially be true of the world in order for the commitment to be
held. For example, in most types of team commitment, we do not expect
participating agents to initially believe that the object of the commitment, rp,
is true. Finally, team commitment is parameterized by a convention c. Team
commitment is then informally defined as follows. A group g is has a team
commitment to goal rp with respect to pre-condition 'I/; and convention c iff:

1. pre-condition 'I/; is initially satisfied; and

2. every agent i E g has an intention of rp until the termination condition is
satisfied;
3. until the termination condition is satisfied, if any agent i E g believes that
the re-evaluation condition of any rule in c is satisfied, then it adopts as an
intention the goal corresponding to the re-evaluation condition, and maintains
this goal until the termination condition is satisfied;

where the termination condition is that one of the goal parts of the convention
rules is satisfied.
More formally, if

is a convention, then we denote the team commitment of group g to goal rp

with respect to convention c and pre-condition 'I/; by (Teame g rp '1/;).
Formally, (Teame g rp '1/;) is defined as follows:
122 Chapter 6

where

Xl � (Int i <p)
and
I

X2 � 1\ (Bel i Pk) => A [(lnt i "(k) U X3]

k=l
and
I

X3 � V"(k.
k=l
We saw in earlier chapters that a single agent can have various different kinds
of commitment towards its goals. We will now see how social variants of these
kinds of commitment can be encoded using our model of conventions and
social commitments.

Blind Social Commitment

Recall from chapter 2 that an agent with a blind commitment towards some
state of affairs <p will maintain <p as an intention until it believes this state
of affairs is achieved. The social convention Cblind corresponding to blind
commitment consists of a single rule.

Cblind� {((Bel i <p), (M-Bel g <p)) }

The re-evaluation condition for this rule is thus that an agent i believes the
goal of the teamwork has been achieved; if it ever comes to believe this, then
it intends that the remainder of the team comes to believe this.
We define the pre-condition for blind commitment as 'ljJblind as follows.

'ljJblind�.(Bel i <p)
In other words, team members do not initially believe that the goal is satisfied.
Expanding out the definition of (Teamblind g <p'ljJ)gives the following.

Vi· (i E g) =>
.(Bel i <p)1\
A ((Int i <p)1\ ((Bel i <p) => A((lnt i xd U xd ) U xd
where

Xl � (M-Bel g <p).
Collective Mental States 123

Suppose an agent working in a team with such a commitment comes to believe

that the goalcp is achieved. Then it must satisfy the following condition:

A((lnt i (M-Bel gcp) ) U (M-Bel gcp) )

Thus i should intend that g mutually believe cp until this state of affairs is
achieved.

A Minimal Social Convention

Next, we specify a minimal social convention that is similar to the Levesque­

Cohen model of joint persistent goals (JPGs) [133, p.98]. A collective with
such a commitment will have a mental state in which:

• initially, every agent does not believe that the goalcp is satisfied, but believes
cp is possible;
• every agent i then has a goal of cp until the termination condition is satisfied
(see below);
• until the termination condition is satisfied, then:
-if any agent i believes that the goal is achieved, then it will have a
goal that this becomes a mutual belief, and will retain this goal until the
termination condition is satisfied;
-if any agent i believes that the goal is impossible, then it will have a
goal that this becomes a mutual belief, and will retain this goal until the
termination condition is satisfied;
• the termination condition is that it is mutually believed that either:
-the goal cp is satisfied; or
-the goal cp is impossible to achieve.

Let

'l/J soc � ....,(Bel icp) A (Bel i EOcp)

and
((Bel i cp) , (M-Bel gcp) ) ,

{ }
'-v--' '-v-"
PI ,1
C soc �
((Bel i A D....,cp) , (M-Bel g A D....,cp) )
'-----v---- ' v
,
P2 ,2
124 Chapter 6

It is not difficult to see that (Teamsocg cp 'l/J soc )expands to:

[(
Vi· (i E g) ::::}
...,(Bel i cp)/\ (Bel i E<)cp)/\
(Int i cp)/\
( ((Bel i cp) ::::} A((lnt i (M-Bel g cp))U xd ) /\
A
((Bel i A D...,cp )::::} A((lnt i (M-Bel g A D...,cp ))U xd )

where

Xl � [(M-Bel g cp)V (M-Bel g A D...,cp )].

6.3 Notes and Further Reading

The fact that mutual belief (or common knowledge) can never arise in a
communication scenario where message delivery is not guaranteed is one of
the most famous results in the modal logic of knowledge [89]. See [54, pp.175-
232] for proofs of this and related results, and [54, pp.230-232] for historical
notes on the result and related work.
The formal treatment of mutual beliefs, desires, and intentions presented in
this chapter is closely based on the presentation of common knowledge in [54].
In particular, the semantics and fixed-point treatments of mutual mental states
are very closely based on those of [54]. Chapters 6 and 11 of [54] are more or
less completely devoted to common knowledge and the associated coordinated
attack problem.
The distinction between mutually intending some state of affairs and the
kinds of mental state involved in teamwork have been noted by many re­
searchers. Searle, in particular, makes a distinction between a group of agents
that are engaged in coordinated activity, but not teamwork (such as a group of
agents driving on a highway), and agents that are actively engaged in team ac­
tion (such a group of ballet dancers) [203]. Early attempts to formalize "team"
mental states include [181, 230].
Following Levesque, Cohen, and Nunes' formalization of teamwork [133],
many attempts to formalize joint intentions appeared in the literature [118,
190]. Jennings demonstrated that such models could be usefully applied to
industrial coordination problems [104, 105, 106].
7 Communication

In this chapter, I show how CORA can be used to capture communication

between agents. Communication has long been recognized as a topic of central
importance in computer science, and many formalisms have been developed
for representing the properties of communicating concurrent systems [98,
154]. Such formalisms have tended to focus on a number of key issues that
arise when dealing with systems that can interact with one another.
Perhaps the characteristic problem in communicating concurrent systems
research is that of synchronizing multiple processes, which was widely studied
throughout the 1970s and 1980s [ 1 1]. Essentially, two processes (cf. agents)
need to be synchronized if there is a possibility that they can interfere with
one another in a destructive way. The classic example of such interference is
the "lost update" scenario. In this scenario, we have two processes, P1 and
P2, both of which have access to some shared variable v. Process P1 begins to
update the value of v, by first reading it, then modifying it (perhaps by simply
incrementing the value that it obtained), and finally saving this updated value in
v. But between P1 reading and again saving the value of v, process P2 updates

v, by saving some value in it. Whenp1 saves its modified value of v, the update

performed by P2 is thus lost, which is almost certainly not what was intended.
The lost update problem is a very real issue in the design of programs that
communicate through shared data structures.
Synchronization is, of course, still an issue when constructing multiagent
systems of the kind under consideration in this book: we must ensure that
agents do not interact with one another in damaging ways [ 194]. However,
such low-level communication problems are generally regarded as solved for
the purposes of multiagent systems research. We usually assume the presence
of some underlying communication platform, which provides appropriate fa­
cilities for robust, synchronized communication between agents. This seems
to be a reasonable assumption: after all, the JAVA programming language pro­
vides exactly such mechanisms.
So, if we do not treat communication in such a "low-level" way, then how
is communication treated by the agent community? In order to understand the
answer, it is helpful to first consider the way that communication is treated
in the object-oriented programming community, that is, communication as
method invocation. Suppose we have a JAVA system containing two objects,
01 and 02, and that 01 has a publicly available method m1. Object 02 can
12 6 Chapter 7

communicate with 01 by invoking method m1. In JAVA, this would mean 02

executing an instruction that looks something like 01. m1 (arg) , where arg is
the argument that 02 wants to communicate to 01. But consider: Which object
makes the decision about the execution of method m1 - is it object 01 or
object 02? In this scenario, object 01 has no control over the execution of m1:
the decision about whether to execute m1 lies entirely with 02.
Now consider a similar scenario, but in an agent-oriented setting. We have
two agents i and j, where i has the capability to perform action a, which
corresponds loosely to a method. But there is no concept in the agent-oriented
world of agentj "invoking a method" on i. This is because i is an autonomous
agent: it has control over both its state and its behavior. It cannot be taken for
granted that agent i will execute action a just because another agent j wants
it to. Performing the action a may not be in the best interests of agent i. The
locus of control with respect to the decision about whether to execute an action
is thus very different in agent and object systems.
In general, agents can neither force other agents to perform some action,
nor write data onto the internal state of other agents. This does not mean
they cannot communicate, however. What they can do is perform actions -
communicative actions - in an attempt to influence other agents appropriately.
For example, suppose I say to you "It is raining in London," in a sincere way.
Under normal circumstances, such a communication action is an attempt by
me to modify your beliefs. Specifically, it is an attempt by me to bring about a
state of affairs that we can characterize in CORA as follows.

(Bel you Weather(/ondon, raining))

Of course, simply uttering the sentence "It is raining in London" is not usu­
ally enough to bring about this state of affairs, for all the reasons that were
discussed above. You have control over your own beliefs (desires, intentions).
You may believe that I am notoriously unreliable on the subject of the weather,
or even that I am a pathological liar. But in performing the communication
action of uttering "It is raining in London," I am attempting to change your
internal state. Furthermore, since this utterance is an action that I perform, I
am performing it for some purpose - presumably because I intend that you

believe it is raining.
Speech act theory treats communication as action. It is predicated on
the assumption that speech actions are performed by agents just like other
actions, in the furtherance of their intentions. Speech act theory is thus broadly
consistent with our view of agents as practical reasoning systems, deciding
Communication 127

moment by moment which action to perform in order to best achieve their

current intentions.
In this chapter, I show how we can model communication as speech acts in
CORA. In later chapters, I demonstrate how these speech acts can be used by
agents in the cooperative problem-solving process and other social actions. I
begin with an historical overview of speech act theory, focusing in particular on
attempts to develop formal theories of speech acts, where communications are
modeled as actions that alter the mental state of communication participants.

7.1 Speech Acts

The theory of speech acts is generally recognized to have begun with the
work of the philosopher John Austin [6]. He noted that a certain class of
natural language utterances - hereafter referred to as speech acts - had
the characteristics of actions, in the sense that they change the state of the
world in a way analogous to physical actions. It may seem strange to think of
utterances changing the world in the way that physical actions do. If I pick up
a block from a table (to use an overworked but traditional example), then the
world has changed in an obvious way. But how does speech change the world?
Austin gave as paradigm examples declaring war and saying "I now pronounce
you man and wife." Stated in the appropriate circumstances, these utterances
clearly change the state of the world in a very tangible way.!
Austin identified a number of performative verbs, which correspond to
various different types of speech acts. Examples of such performative verbs are
request, iriform, and promise. In addition, Austin distinguished three different
aspects of speech acts: the locutionary act, or act of making an utterance (e.g.,
saying "Please make some tea"), the illocutionary act, or action performed in
saying something (e.g., "He requested me to make some tea"), and perlocution,
or effect of the act (e.g., "He got me to make tea").
Austin referred to the conditions required for the successful completion
of performatives asfelicity conditions. He recognized three important felicity
conditions:

1. There must be an accepted conventional procedure for the performative,

1 Notice that when referring to the effects of communication, I am ignoring "pathological"

cases, such as shouting while on a ski run and causing an avalanche. Similarly, I will ignore
"microscopic" effects (such as the minute changes in pressure or temperature in a room caused by
speaking).
12 8 Chapter 7

and the circumstances and persons must be as specified in the procedure.

2. The procedure must be executed correctly and completely.
3. The act must be sincere, and any uptake required must be completed, insofar
as is possible.

Austin's work was extended by John Searle, in his 1969 book Speech
Acts [202]. Searle identified several properties that must hold for a speech act
performed between a hearer and a speaker to succeed. For example, consider
a request by SPEAKER to HEARER to perform ACTION:

1. Normal I/O conditions. Normal liD conditions state that HEARER is able to
hear the request (thus must not be deaf, . . . ), the act was performed in normal
circumstances (not in a film or play, . . . ), etc.
2. Preparatory conditions. The preparatory conditions state what must be true
of the world in order that SPEAKER correctly choose the speech act. In this
case, HEARER must be able to perform ACTION, and SPEAKER must believe
that HEARER is able to perform ACTION. Also, it must not be obvious that
HEARER will do ACTION anyway.

3. Sincerity conditions. These conditions distinguish sincere performances of

the request; an insincere performance of the act might occur if SPEAKER did
not really want ACTION to be performed.

Searle also attempted a systematic classification of possible types of speech

acts, identifying the following five key classes:

1. Representatives. A representative act commits the speaker to the truth of an

expressed proposition. The paradigm case is informing.
2. Directives. A directive is an attempt on the part of the speaker to get the
hearer to do something. Paradigm case: requesting.
3. Commissives. Commit the speaker to a course of action. Paradigm case:
promising.
4. Expressives. Express some psychological state (e.g., gratitude). Paradigm
case: thanking.
5. Declarations. Effect some changes in an institutional state of affairs.
Paradigm case: declaring war.

In the late 1960s and early 1970s, a number of researchers in AI began to build
systems that could plan how to autonomously achieve goals [2]. Clearly, if
such a system is required to interact with humans or other autonomous agents,
Communication 12 9

then such plans must include speech actions. This introduced the question of
how the properties of speech acts could be represented such that planning
systems could reason about them. Cohen and Perrault [39] gave an account
of the semantics of speech acts by using techniques developed in AI planning
research [59]. The aim of their work was to develop a theory of speech acts

. . . by modeling them in a planning system as operators defined . . . in terms of speakers'

and hearers' beliefs and goals. Thus speech acts are treated in the same way as physical
actions. [39]

The formalism chosen by Cohen and Perrault was the STRIPS notation,
in which the properties of an action are characterized via pre- and post­
conditions [59]. The idea is very similar to Hoare logic [97]. Cohen and
Perrault demonstrated how the pre- and post-conditions of speech acts such
as request could be represented in a multimodal logic containing operators for
describing the beliefs, abilities, and wants of the participants in the speech act.
Consider the Request act. The aim of the Request act will be for a speaker
to get a hearer to perform some action. Figure 7. 1 defines the Request act. Two
preconditions are stated: the "cando.pr" (can-do pre-conditions), and "want.pr"
(want pre-conditions). The cando.pr states that for the successful completion of
the Request, two conditions must hold. First, the speaker must believe that the
hearer of the Request is able to perform the action. Second, the speaker must
believe that the hearer also believes it has the ability to perform the action.
The want.pr states that in order for the Request to be successful, the speaker
must also believe it actually wants the Request to be performed. If the pre­
conditions of the Request are fulfilled, then the Request will be successful: the
result (defined by the "effect" part of the definition) will be that the hearer
believes the speaker believes it wants some action to be performed.
While the successful completion of the Request ensures that the hearer
is aware of the speaker's desires, it is not enough in itself to guarantee that
the desired action is actually performed. This is because the definition of
Request only models the illocutionary force of the act. It says nothing of the
perlocutionary force. What is required is a mediating act. Figure 7. 1 gives
a definition of CauseToWant, which is an example of such an act. By this
definition, an agent will come to believe it wants to do something if it believes
that another agent believes it wants to do it. This definition could clearly be
extended by adding more pre-conditions, perhaps to do with beliefs about
social relationships, power structures, etc.
The Inform act is as basic as Request. The aim of performing an Inform
130 Chapter 7

Request(S,H,a)
Preconditions Cando.pr (S BELIEVE (H CANDOa)) 1\
(S BELIEVE (H BELIEVE (H CANDOa)))
Want.pr (S BELIEVE (S WANT requestlnstance))
Effect (H BELIEVE (S BELIEVE (S WANTa)))

CauseToWant(Al ,A2,a)
Preconditions Cando.pr (Al BELIEVE (A2 BELIEVE (A2 WANTa)))
Want.pr x

Effect (Al BELIEVE (Al WANTa))

Injorm(S,H, cp)
Preconditions Cando.pr (S BELIEVE cp)
Want.pr (S BELIEVE (S WANT iriformlnstance))
Effect (H BELIEVE (S BELIEVE cp))

Convince(Al ,A2, cp)

Preconditions Cando.pr (Al BELIEVE (A2 BELIEVE cp))
Want.pr x

Effect (Al BELIEVE cp)

Figure 7.1
Definitions from Cohen and Perrault's plan-based theory of speech acts.

will be for a speaker to get a hearer to believe some statement. Like Request,
the definition of Inform requires an associated mediating act to model the
perlocutionary force of the act. The cando.pr of Inform states that the speaker
must believe cp is true. The effect of the act will simply be to make the hearer
believe that the speaker believes cpo The cando.pr of Convince simply states
that the hearer must believe that the speaker believes cpo The effect is simply to
make the hearer believe cpo

Speech Acts as Rational Action

While the plan-based theory of speech acts was a major step forward, it was
recognized that a theory of speech acts should be rooted in a more general
theory of rational action. This observation led Cohen and Levesque to develop
a theory in which speech acts were modeled as actions performed by rational
agents in the furtherance of their intentions [36]. The foundation upon which
they built this model of rational action was their theory of intention, described
in [35]. The formal theory is too complex to describe here, but as a flavor, here
is the Cohen-Levesque definition of requesting, paraphrased in English:
Communication 131

A request is an attempt on the part of spkr, by doing e, to bring about a state where,
ideally, (i) addr intends a, (relative to the spkr still having that goal, and addr still being
helpfully inclined to spkr), and (ii) addr actually eventually does a, or at least brings
about a state where addr believes it is mutually believed that it wants the ideal situation.
[36, p. 241]

I use the Cohen-Levesque model of speech acts as the basis on which to

develop the CORA model of communication later in this chapter.

Agent Communication Languages: KQML and FIPA

Throughout the 1980s and 1990s, interest in multiagent systems developed

rapidly [ 18]. An obvious problem in multiagent systems is how to get agents
to communicate with one another. To this end, in the early 1990s, the DARPA
Knowledge Sharing Effort (KSE) began to develop the Knowledge Query and
Manipulation Language (KQML) and the associated Knowledge Interchange
Format (KIF) as a common framework via which multiple expert systems (cf.
agents) could exchange knowledge [ 170, 147].
KQML is essentially an "outer" language for messages: it defines a simple
LISP-like format for messages, and 4 1 performatives, or message types, that
define the intended meaning of a message. Example KQML performatives
include ask-if and te11. The content of messages was not considered part
of the KQML standard, but KIF was defined to express such content. KIF is
essentially classical first-order predicate logic, recast in a LISP-like syntax.
To better understand the KQML language, consider the following example
(from [ 147, p.354]):

(ask-one

:content (PRICE IBM ?price)

:receiver stock-server

:language LPROLOG

:ontology NYSE-TICKS

The intuitive interpretation of this message is that the sender is asking about the
price of IBM stock. The performative is ask-one, which an agent will use to
ask a question of another agent where exactly one reply is needed. The various
other components of this message represent its attributes. The most important
of these is the :content field, which specifies the message content. In this
case, the content simply asks for the price of IBM shares. The :receiver at­
tribute specifies the intended recipient of the message, the: language attribute
132 Chapter 7

specifies that the language in which the content is expressed is called LPROLOG
(the recipient is assumed to "understand" LPROLOG) , and the final: ontology
attribute defines the terminology used in the message.
Formal definitions of the syntax of KQML and KIF were developed by the
KSE, but KQML lacked any formal semantics until Labrou and Finin's [ 127].
These semantics were presented using a pre- and post-condition notation,
closely related to Cohen and Perrault's plan-based theory of speech acts [39].
These pre- and post-conditions were specified by Labrou and Finin using a
logical language containing modalities for belief, knowledge, wanting, and
intending.
The take-up of KQML by the multiagent systems community was signif­
icant. However, Cohen and Levesque (among others) criticized KQML on a
number of grounds [38], the most important of which being that the language
was missing an entire class of performatives - commissives, by which one
agent makes a commitment to another. As Cohen and Levesque point out, it is
difficult to see how many multiagent scenarios could be implemented without
commissives, which appear to be important if agents are to coordinate their
actions with one another.
In 1995, the Foundation for Intelligent Physical Agents (FIPA) began its
work on developing standards for agent systems. The centerpiece of this initia­
tive was the development of an ACL [60]. This ACL is superficially similar to
KQML: it defines an "outer" language for messages, it defines 20 performatives
(such as inform) for defining the intended interpretation of messages, and it
does not mandate any specific language for message content. In addition, the
concrete syntax for FIPA ACL messages closely resembles that of KQML. Here
is an example of a FIPA ACL message (from [60, p.lO]):

(inform

: sender agent1

:receiver agent2

:content (price good2 150)

:language sl

:ontology hpl-auction

The FIPA communication language is similar to KQML; the relationship is dis­

cussed in [60, pp.68-69]. The 20 performatives provided by the FIPA commu­
nication language are categorized in Table 7. 1.
Communication 133

Table 7.1
Performatives provided by the FIPA communication language.

Passing Requesting Negotiation Performing Error

Performative Information Information Actions Handling
accept-proposal X
agree X
cancel X X
cfp X
confirm X
disconfirm X
failure X
inform X
inform-if X
inform-ref X
not-understood X
propose X
query-if X
query-ref X
refuse X
reject-proposal X
request X
request-when X
request-whenever X
subscribe X

The FIPA ACL has been given a formal semantics, in terms of a Semantic
Language (SL). The approach adopted for defining these semantics draws
heavily on [36], but in particular on Sadek's enhancements to this work [23]. SL
is a quantified multimodal logic, which contains modal operators for referring
to the beliefs, desires, and uncertain beliefs of agents, as well as a simple
dynamic logic-style apparatus for representing agent's actions. The semantics
of the FIPA ACL map each ACL message to a formula of SL, which defines a
constraint that the sender of the message must satisfy if it is to be considered
as conforming to the FIPA ACL standard. FIPA refers to this constraint as the
feasibility condition. The semantics also map each message to an sL-formula
that defines the rational effect of the action - the "purpose" of the message:
what an agent will be attempting to achieve in sending the message (cf.
perlocutionary act). However, in a society of autonomous agents, the rational
effect of a message cannot (and should not) be guaranteed. Hence conformance
does not require the recipient of a message to respect the rational effect part of
the ACL semantics - only the feasibility condition.
To illustrate the FIPA approach, I give an example of the semantics of the
134 Chapter 7

FIPA inform performative [60, p.25]:

(i, inform(j, <p))

feasibility pre-condition: B;<p 1\ -,Bi(Bijj<p V Uijj<p) (7. 1)
rational effect: Bj<p

The B; is a modal connective for referring to the beliefs of agents (rather like
Bel in .cORA); Bif is a modal connective that allows us to express whether
an agent has a definite opinion one way or the other about the truth or falsity
of its parameter; and Uif is a modal connective that allows us to represent the
fact that an agent is "uncertain" about its parameter. Thus an agent i sending
an inform message with content <p to agent) will be respecting the semantics
of the FIPA ACL if it believes <p, and it is not the case that it believes of) either
that) believes whether <p is true or false, or that) is uncertain of the truth or
falsity of <po
In the remainder of this chapter, I use ideas from the work of Cohen and
Perrault [39], Cohen and Levesque [36], Bretier and Sadek [23], and from the
FIPA standard itself [60] to define the semantics of "inform" (representative)

and "request" (directive) speech acts. In the style of [60], I then show how
these two basic speech acts can be used to define a range of other possible
speech acts. I begin by defining what it means for one agent to attempt to bring
about some state of affairs.

7.2 Attempts

We will see below that when agents communicate with one another, they are in
fact attempting to bring about some state of affairs. The notion of an attempt
will thus be a central component of our formalization of communication. The
model of attempts I adopt is based that proposed by Cohen and Levesque [36,
p.240]. The idea is that an attempt by agent i to bring about a state <p is an
action a, which is performed by i with the desire that after a is performed, <p
is satisfied, but with the intention that at least 'l/J is satisfied. The ultimate aim
of the attempt - the thing that i hopes to bring about - is represented by <p,
whereas 'l/J represents "what it takes to make an honest effort" [36, p.240]. If i
is successful, then bringing about 'l/J will be sufficient to cause <po To make this
concrete, consider the following scenario.

Alice, an academic, wishes to obtain tenure. Unfortunately, she is in competition for

promotion with another colleague, Bob. She has no plan available to her that she
Communication 135

believes will guarantee to bring about promotion. However, if the tenure committee
could somehow be convinced that Bob had been faking research results, then this may
be sufficient to scupper Bob's chances. To this end, she sends a letter to the tenure
committee stating that Bob had been faking his results.

In this scenario, Alice is performing an action (sending the letter to the tenure
committee) in an attempt to gain promotion. Alice intends that this action will
cause the tenure committee to believe that Bob is a cheat, and she desires that
this will be enough to foil Bob's chances. Whether or not it is enough will be
outside Alice's control.
Formally, an attempt by i to achieve rp by performing 0:, at least achieving
'ljJ, is written {Attempt i 0: rp 'ljJ}; following Cohen and Levesque, I use
curly brackets here to indicate that attempts are complex actions, rather than
predicates or modal operators [36, p.240].

[ (Bel hrp) /\ 1
(Agt 0: i) /\
{Attempt io: rp'ljJ} � ?; 0:
(Des i (Achvs 0: rp)) /\
(Int i (Achvs 0:'ljJ))

Returning to the scenario above, the action 0: is Alice sending the letter about
Bob faking his research results; the state of affairs rp that Alice wants to bring
about is being promoted, and'ljJ, what it takes to make a reasonable effort, is
the tenure committee believing Bob is a cheat.
We now prove some properties of attempts. First, if an agent attempts to
bring about some state of affairs rp by at least bringing about'ljJ, then it desires
that rp is at least possible.

THEOREM 7. 1: FP (Happens {Attempt i 0: rp'ljJ}) => (Des i EOrp)

Proof Assume (M, V, w, p) FP (Happens {Attempt i 0: rp 'ljJ}) for arbi­

trary (M, V, w, p). Then by expanding out the definition of Attempt, we have
(M, V, w, p) FP (Happens (Des i (Achvs 0: rp))?) and so (M, V, w, p(O)) FP
(Des i (Achvs 0: rp)). Now, since FS (Achvs 0: rp) => EOrp, then from the K
axiom for desires it is easy to obtain (M, V, w, p(O)) FP (Des i EOrp)) and
we are done. _

Similarly, if an agent attempts to bring about some state of affairs rp by at least

bringing about'ljJ, then it believes that'ljJ is possible.
136 Chapter 7

THEOREM 7.2: i=p (Happens {Attemptia cp1/!}) ::::} (Desi E<>1/!)

Proof As Theorem 7. 1. •

7.3 Informing

Probably the simplest form of communication that can take place between
two agents involves the exchange of information. Such information exchange
will usually result in belief changes. Attempts to communicate information are
known as representative speech acts [202]. The paradigm example of repre­
sentative acts involves agents uttering simple declarative sentences. Examples
include an agent saying "It is raining in London," or "Tony Blair is Prime Min­
ister." I will say an agent performing a representative act is iriforming another
agent about some state of affairs.
It is possible to identify the following properties of inform actions that we
will capture in our model.

1. An inform speech act (and other speech acts) takes place between a
"speaker" agent and one or more "hearer" agents.
Most theories of speech acts presuppose that they take place between just two
agents: a "speaker" and a "hearer." The model of speech acts we will see below
is more general than this - it allows for speech acts where the "hearer" is a
group of agents. For example, consider somebody making a presentation to a
group of people. The speaker is attempting to bring about some mutual belief
in the mental state of the hearers.
2. Inform actions are actually actions that an agent may perform, just like any
other action.
The model of speech acts I present here does not presuppose any particular
"transport mechanism" for the communication. For example, it does not re­
quire a FIPA-like communication language for messages. Instead, any action
that satisfies certain properties can be counted as an inform message. For exam­
ple, uttering the sentence "The door is open," writing a note saying "The door
is open," and simply nodding in the direction of the door can all be counted
as attempts by an agent to inform a hearer that the door is open if the actor
satisfies certain properties.
3. The purpose of an inform action is to modify the beliefs of agent(s) hearing
the speech acts, and in particular, to cause them to believe some state of affairs.
The state of affairs that the agent performing the speech act desires to bring
Communication 137

about is known as the perlocutionary force in speech act theory [202], or the
rational effect of a message in FIPA [60]. The rational effect of an inform action
is to cause some particular belief in the hearer.
4. Since the speaker of the inform act cannot normally bring about the ultimate
intention of the speech act in the listener(s), they will attempt to bring about
some "lesser" belief, such that this will be sufficient, if the listener(s) are so
inclined, to cause the listener(s) to believe the ultimate belief.
There is something of a paradox here. If I am completely autonomous, and
exercise complete control over my mental state, then nothing you say will have
any effect on my mental state. But if nothing you say can change my state,
then how can we communicate? The answer I adopt here is that if you are
attempting to inform me of some state of affairs, then the best you can do is
convince me that you believe this state of affairs. If I am then appropriately
inclined, this will be sufficient to cause me to believe it.
5. Inform actions (and other speech acts) are attempts to bring about some
intended state of affairs - they may fail to bring this state of affairs.
A model of speech acts that in some sense required an agent to bring about
the intended change in mental state would be too strong. An agent can attempt
to bring about such a mental state, but may fail for at least two reasons. First,
the action itself may fail - for example, the communications medium may be
faulty. Second, for the reasons discussed above, the hearer(s) of the speech act
may simply choose to ignore it.
6. Inform actions (and other speech acts) cannot be performed accidentally.
To modify an example from [36, p.239], suppose a blindfolded person reaches
into a bowl of flashcards, pulls out three cards, and knowingly turns them
toward another person. The cards say "The door is open." One would not be
inclined to say that an inform speech act took place.

It is normally required that inform actions be performed sincerely: that the

speaker must believe the state of affairs referred to in the act. However, the
model of inform speech acts presented here does not require that informs be
sincere. I will instead focus on the (more general case) of agents that are simply
attempting to bring about a belief in some conversation participant.
We can now formally define the Inform speech act. We write

{Inform i g a 'P}

to indicate that a is an action performed by agent i in an attempt to inform

group g of 'P. Note the use of curly brackets here: Inform is neither a predicate
138 Chapter 7

nor a modal operator, but an action.

The definition of Inform directly builds on that of Attempt. We say agent
i informs group g of <p through performing action a if a is an attempt to bring
about the mutual belief in g of <p, by at least making it mutually believed in g
thati intends that g mutually believe <po Formally, this gives:

{Informi g a <p} � {Attempt i a 'lj; X}

where

'lj; � (M-Bel g<p)

and

X � (M-Bel g (Inti (M-Bel g<p))).

In other words, action a counts an inform action between agents i (speaker)

and group g (hearers) with respect to content <p if a is an attempt by i to cause
the hearers to mutually believe <p, by at the very least causing g to mutually
believe thati intends that g mutually believe <po
We can prove some properties of Inform speech acts. First, we can prove
that if one agent informs a group of <p, then the informant desires the hearers
may come to mutually believe <po

THEOREM 7.3: F'P (Happens {Inform i g a <p}) ::::} (Desi EO(M-Bel g<p))

Proof Assume (M, V, w,p) F'P (Happens {Inform i g a <p}) for arbi­
trary (M, V, w,p). From the definition of Inform, we have (M, V, w,p) F'P
(Happens {Attempt i a 'lj; X}) where

'lj; � (M-Bel g<p)

and

X � (M-Bel g (Inti (M-Bel g<p))).

The desired result then follows immediately from Theorem 7. 1. •

We can also prove that the informant does not believe the hearers already have
a mutual belief of <po Notice that an agent can be informing a group of agents
of <p even though every agent in the group already individually believes <po In
informing the group of <p, the speaker is attempting to raise the status of this
Communication 139

belief to the mutual belief of cp, which is not implied simply by everyone in a
group believing cpo

THEOREM 7.4: FP (Happens {Inform i g a cp}) => ...., (Bel i (M-Bel g cp))

Proof Immediate by expanding the definitions of Inform and Attempt. •

Note that the case of "one-to-one" inform speech acts is covered by the
multiagent definition given above. Where we wish to deal with an agent i
informing a single agent j of cp, we will abuse notation slightly and write
{Inform ij a cp} rather than the more cumbersome {Inform i {j} a cp}.

7.4 Requesting

Inform speech acts are attempts by a speaker to get the hearer to believe some
state of affairs. In contrast, request speech acts (directives) are attempts by a
speaker to modify the intentions of the hearer. However, we can identify at
least two different types of requests:

• Requests to bring about some state of affairs.

An example of such a request would be when one agent said "Keep the door
closed." We call such requests "requests-that."
• Requests to perform some particular action.

An example of such a request would be when one agent said "Lock the door."
We call such requests "requests-to."

Requests-that are more general than requests-to. In the former case (requests­
that), the agent communicates an intended state of affairs, but does not com­
municate the means to achieve this state of affairs. Me asking you to keep the
door closed does not tell you how I want you to keep the door closed - you
may lock it, prop it closed with a chair, or sit against it. In the case of request­
ing to, however, the agent does not communicate the desired state of affairs
at all. Instead, it communicates an action to be performed, and the state of af­
fairs to be achieved lies implicit within the action that was communicated. It is
useful to make a distinction between the two cases, and for this reason, I will
formalize both. (As it turns out, it is straightforward to encode requests-to as a
special case of requests-that.)
Just as we identified desirable properties of inform speech acts, so it is
possible to identify the following properties of request actions.
140 Chapter 7

1. A request speech act takes place between a "speaker" agent and one or more
"hearer" agents.
2. Request actions are actually actions that an agent may perform, just like
any other action.
3. The purpose of a request action is to modify the intentions of agent(s)
hearing the speech act, and in particular, to cause them to intend some state
of affairs (in the case of requesting-that), or to intend to perform some action
(in the case of requesting-to).
4. Since the speaker of the request act cannot normally bring about the ulti­
mate intention of the speech act directly in the listener(s), they will attempt to
bring about a "lesser" belief, such that this will be sufficient, if the listener(s)
are so inclined, to cause the listener(s) to intend the ultimate state of affairs (in
the case of requesting-that), or to intend to perform the action (requesting-to).
5. Request actions are attempts to bring about some intended state of affairs
- they may fail to bring this state of affairs.
6. Request actions cannot be performed accidentally.

Given this discussion, it is easy to formally define the RequestTh speech act. I
write

{RequestTh iga cp}

to indicate that a is an action performed by agent i in an attempt to request

that group g intends cpo Note again the use of curly brackets to indicate that
RequestTh is neither a predicate nor a modal operator, but an action.
The definition of RequestTh is as follows.

{RequestTh ig a cp} � {Attempt ia'ljJ X}

where

'ljJ � (M-Intg cp)

and

X � (M-Bel g (Int i (M-Intg cp))).

We can prove that the speaker does not believe the hearer already intends cpo

THEOREM 7 .5: FP (Happens {RequestTh iga cp}) => ...., (Bel i (M-Intg cp))
Communication 141

Proof Immediate from the definitions of Req uestTh and Attempt. •

We can define the request-to speech act as a special case of requesting that. A
request to perform a is thus simply a request that a is performed.

{RequestToi g a a' l � {RequestTh i g a A<> (Happens a/ n .

7.5 Composite Speech Acts

Inform and request speech acts are the basic mechanisms via which agents
will communicate: they define the possible ways in which agents can attempt
to modify one another's beliefs and intentions - their key mental states. In
this section, we will see how a number of different speech acts can be defined
in terms of inform and request acts. In particular, I will define some simple
conversation policies that make use of these primitives:
Language inter-operability requires more than simply that the agents agree on the
format and meaning of the various primitive [communication language] messages.
As a practical matter, agents must also agree on the range of possible sequences and
contexts of messages when they are interpreted in the context of larger [... ] dialogues,
or conversations. [8 3, p.1]

I make use of a simple graphical state-machine notation for specifying con­

versation policies. In this notation, nodes in the conversation policy graph cor­
respond to states of the conversation, and arcs to the performance of speech
acts. A node with an arrow entering it, but no predecessor node, is the start
state. A node marked by a double circle border is an end state. I will start by
considering question and answer conversations.

Questions and Answers

We have already seen that the Inform speech act is the basic mechanism via
which agents can exchange information. But consider: Why should one agent
inform another of anything? There are a number of possibilities why agenti
should inform agentj of rp, for example:

1. Agent i has an intention that agent j believe rp for its own reasons. For
example, returning to the Alice and Bob tenure example from earlier in this
chapter, Alice may intend that the tenure committee believe that Bob faked
research results because this will enable her intention of gaining tenure to be
satisfied.
142 Chapter 7

Figure 7.2
The conversation policy associated with Query.

2. Agent) has explicitly asked agenti the truth status of r.p.

The first case is simply an instance of the more general situation, where an
agent performs an action because the agent believes it is in its best interests.
We will therefore focus on the second case, where one agent explicitly asks
another about the truth status of some proposition.
We can easily define a Query speech act. The idea is that an agent i will
perform a Query action if it has an intention of knowing the truth status of some
statement r.p. For example, I may want to know whether or not it is raining, and
if I believe that you know whether or not it is raining, I can query you to this
effect. Formally, we write {Queryi) r.p} to denote a query action performed by
a speaker i to a hearer) with the intention of discovering whether or not r.p is
the case. Query is defined, in the terminology of FIPA, as a macro action:

{Queryi) Q r.p} � {RequestTh i) Q (Bel i r.p) V (Bel hr.p)}

In other words, a Query by agenti to) with respect to statement r.p is a request
fromi to) to bring about the belief ini that either r.p is true or that r.p is false.
The mechanism via which agent) can bring about such belief is the Inform
speech act. This leads to the conversation policy illustrated in Figure 7.2.

Request-based Conversation Policies

When I ask you to make a cup of tea, the conversation does not usually end
there. Whether or not you chose to accede to my request is up to you, but
hopefully you will either confirm to me that you will carry out the request, or
else explicitly tell me that you will not. To define this kind of dialogue, we will
make use of two auxiliary definitions. The first, Agree, will be used to capture
Communication 143

{Refuse j i a3 <X2}

Figure 7.3
The conversation policy associated with Req uest To.

the case where you explicitly state that you intend to perform some action.

{Agreei) a a' } � {Informi) a (Agt a' i) 1\ (Inti AO(Happens a' ))}

Thus agent i agreeing to perform an action a' to agent) simply means that i
informs) that it is the agent of a' , and that it intends that a' is performed. (Note
that a is the action of informing, rather than the action being referred to in the
inform.)
The Refuse speech act is used to capture the situation where i is telling)
that it will not carry out an action.

{Refusei) a a' } � {Informi) a -,(Inti AO(Happens a' ))}

Using these two speech acts, we can define a conversation policy for Request
speech acts. The policy is illustrated in Figure 7.3.

7.6 Notes and Further Reading

The problems associated with communicating concurrent systems have driven

a significant fraction of research into theoretical computer science since the
early 1980s. Two of the best-known formalisms developed in this period are
Tony Hoare's Communicating Sequential Processes (csps) [98], and Robin
Milner's Calculus of Communicating Systems (ccs) [ 154]. Temporal logic
has also been widely used for reasoning about concurrent systems - see, for
example, Pnueli [ 177] for an overview. A good reference, that describes the
key problems in concurrent and distributed systems, is Ben-Ari [ 1 1].
Problems such as synchronization were widely studied in the early days
of multiagent systems research, though they are less in the mainstream of re­
search today. Georgeff developed formalisms for representing multiagent ac-
144 Chapter 7

tivities, and algorithms for synchronizing multiagent plans that drew upon syn­
chronization techniques developed in theoretical computer science [75, 76].
Rosenschein developed similar techniques, which combined work on speech
acts with eSP-like communication to synchronize multiagent plans [ 194]. Stu­
art showed how multiagent plans could be encoded as formulae of temporal
logic: showing the plans were interference-free then reduced to a problem of
showing that the resulting formula was satisfiable, a well-understood (if some­
what complex) logic problem [236].
The plan-based theory of speech acts developed by Cohen and Perrault
made speech act theory accessible and directly usable to the artificial intel­
ligence community [39]. In the multiagent systems community, this work is
arguably the most influential single publication on the topic of speech act-like
communication. Many authors have built on its basic ideas. For example, bor­
rowing a formalism for representing the mental state of agents that was devel­
oped by Robert Moore [ 157], Douglas Appelt was able to implement a system
that was capable of planning to perform speech acts [3, 4].
Many other approaches to speech act semantics have appeared in the liter­
ature. For example, Perrault [ 174] described how Reiter's default logic [ 193]
could be used to reason about speech acts. Appelt gave a critique of Per­
rault's work [5, pp.l67-168], and Konolige proposed a related technique us­
ing hierarchic auto-epistemic logic (HAEL) [ 12 1] for reasoning about speech
acts. Galliers emphasized the links between speech acts and AMG belief
revision [7 1]: she noted that the changes in a hearer's state caused by a
speech act could be understood as analogous to an agent revising its be­
liefs in the presence of new information [69]. Singh developed a theory of
speech acts [2 14, 2 15] using his formal framework for representing ratio­
nal agents [2 1 1, 2 10, 2 12, 2 18, 2 13, 2 16, 2 17]. He introduced a predicate
comm(i,), m) to represent the fact that agent i communicates message m to
agent), and then used this predicate to define the semantics of assertive, direc­
tive, commissive, and permissive speech acts.
At the time of writing (Spring 2000), there is a great deal of industrial in­
terest in agent technology. A lot of this interest is focussed on the development
of standardized agent communication languages, and in particular, on the FIPA
standardization effort [67]. As noted in the text, the FIPA communication lan­
guage does have a formally defined semantics, based largely on the techniques
of Cohen and Levesque [36] and Sadek et al. [23]. One key issue for this work
is that of semantic conformance testing. The conformance testing problem can
be summarized as follows [241]: We are given program 7f;, and an agent com-
Communication 145

munication language Le with the semantics [ . . .]Lc- The aim is to determine

whether or not Wi respects the semantics [ . . .]Lc whenever it communicates
using Le. (Syntactic conformance testing is of course trivial.)
The importance of conformance testing has been recognized by the ACL
community [60, p.l]. However, to date, little research has been carried out ei­
ther on how verifiable communication languages might be developed, or on
how existing ACLs might be verified. One exception is my [241], where the is­
sue of conformance testing is discussed from a formal point of view: I point out
that ACL semantics are generally developed in such a way as to express con­
straints on the senders of messages. For example, the constraint imposed by
the semantics of an "inform" message might state that the sender believes the
message content. This constraint can be viewed as a specification. Verifying
that an agent respects the semantics of the agent communication language then
reduces to a conventional program verification problem: show that the agent
sending the message satisfies the specification given by the communication
language semantics. In [24 1] I pointed out that this poses the following prob­
lem for ACL conformance testing. The formalisms used to give a semantics to
communication languages are typically quantified multimodal logics, similar
to CORA. However, we have no way of systematically assigning mental states
to computer programs, and so we cannot verify whether or not such programs
implement "mentalistic" semantics. In the CORA case, for example, we have
no procedure that will take an arbitrary program and from that program gener­
ate a CORA model that represents the properties of the program. This makes
it (at best) very doubtful whether one could realistically hope to do semantic
conformance testing for languages with FIPA-like semantics. Recently, a num­
ber of alternative proposals have appeared for communication languages with
a verifiable semantics [209, 175, 243]. See [ 128] for a discussion of the state
of the art in agent communication languages as of early 1999.
8 Cooperation

Agents - both human and artificial - can engage in many and varied types of
social interaction. Communicative acts, such as informing and requesting,are
probably the simplest. In this chapter, however, we will consider cooperative
problem solving, a much more sophisticated and structured form of social
interaction. Cooperative problem solving occurs when a group of autonomous
agents choose to work together to achieve a common goal. Note that we are
not interested in cases where the agents working together are not autonomous.
Client-server computing is an obvious example of non-cooperative problem
solving,as are all "master-slave" systems [194].
Examples of cooperative problem solving in everyday life include a group
of people working together to move a heavy object, play a symphony,build a
house, or write a joint paper. In this chapter, I show how CORA can be used
to construct a formal model of such cooperative problem solving. I begin by
informally discussing what cooperative problem solving is.

8.1 What Is Cooperative Problem Solving?

Comparatively few of our intentions can be achieved without the cooperative

participation of our peers. Even intentions that at first sight appear to involve
just ourselves,such as traveling to an airport,require cooperation on the part of
other agents on closer inspection. One of the main questions in the social sci­
ences, and increasingly in the multiagent systems community, is how and why
autonomous agents should cooperate with one another. As Castelfranchi points
out,cooperation seems to be something of a paradox because it apparently in­
volves giving up some of our autonomy [28]. In agreeing to work with you, I
am effectively making my own intentions (and potentially my beliefs and de­
sires) subservient to yours for the duration of our cooperative act. Paradox or
not, cooperative action, and in particular, cooperative problem solving, does
occur. In this chapter, I will present a model that attempts to account for the
mental state of agents as they are engaged in the cooperative problem-solving
process.
148 Chapter 8

The model consists of four stages:

1. Recognition
The cooperative problem solving process begins when some agent recognizes
the potential for cooperative action. This recognition may come about because
an agent has a goal that it does not have the ability to achieve on its own, or
else because the agent prefers a cooperative solution.
2. Team formation
During this stage,the agent that recognized the potential for cooperative action
at stage (1) solicits assistance. If this stage is successful,then it will end with a
group of agents having some kind of nominal commitment to collective action.
3. Plan formation
During this stage,the agents attempt to negotiate a joint plan that they believe
will achieve the desired goal.
4. Team action
During this stage, the newly agreed plan of joint action is executed by the
agents, which maintain a close-knit relationship throughout. This relationship
is defined by a convention, which every agent follows.

The key theme in the model is that as cooperative problem solving progresses,
the group of participating agents build up increasingly greater commitment to
the process. When the process begins, there will in fact be no kind of group
commitment at all - cooperation will "exist" only in the mental state of the
agent that initiates the process. However,as the process proceeds,the team will
build up first a commitment to the principle of collective action, and then to
the practice of it.
A corollary of the fact that agents are autonomous is that cooperation may
fail. If agents are not required to cooperate, then sometimes they will not.
Even when initial cooperation is established,it can subsequently fail for many
different reasons. For example, a group of agents that agree to cooperate in
principle may discover that the assumptions upon which their choices were
made do not in fact hold. Alternatively,events beyond the control of the team
may make successful completion of their cooperation impossible. An adequate
theory of cooperation must recognize that such failure is possible,identify the
key points at which it may occur, and characterize the behavior of a rational
agent in such circumstances.
Another aspect of the model is that it assumes communication is the
main tool through which collective mental states will be assembled. Although
Cooperation 149

something resembling cooperation is possible without communication [225],

almost all forms of cooperative problem solving in the real world rely upon
explicit,high-level communication in some form or other.
Note that in reality, the four stages are iterative, in that if one stage fails,
the agents may return to previous stages.

8.2 Recognition

Cooperative problem solving begins when some agent in a multiagent commu­

nity has a goal,and recognizes the potential for cooperative action with respect
to that goal. Recognition may occur for several reasons. The paradigm case is
that in which the agent is unable to achieve the goal in isolation, but believes
that cooperative action can achieve it. For example,an agent may have a goal
which,to achieve,requires information that is only accessible to another agent.
Without the cooperation of this other agent,the goal cannot be achieved. More
prosaically,an agent with a goal to move a heavy object might simply not have
the strength to do this alone.
Alternatively,an agent may be able to achieve the goal on its own,but may
not want to. There may be several reasons for this. First, it may believe that
in working alone, it will clobber one of its other goals. For example, suppose
I have a goal of lifting a heavy object. I may have the capability of lifting the
object, but I might believe that in so doing, I would injure my back, thereby
clobbering my goal of being healthy. In this case, a cooperative solution -
involving no injury to my back - is preferable. More generally,an agent may
believe that a cooperative solution will in some way be better than a solution
achieved by action in isolation. For example,a solution might be obtained more
quickly,or may be more accurate as a result of cooperative action.
Believing that you either cannot achieve your goal in isolation,or that (for
whatever reason) you would prefer not to work alone, is part of the potential
for cooperation. But it is not enough in itself to initiate the social process. For
there to be potential for cooperation with respect to an agent's goal, the agent
must also believe there is some group of agents that can actually achieve the
goal; see Figure 8.1.

Ability

In order to precisely define the conditions that characterize the potential for
cooperative action, it is necessary to introduce a number of subsidiary def-
150 Chapter 8

Figure 8.1
Agent A 1 recognizes the potential for cooperative action.

initions. First, we require definitions of single-agent and multiagent ability:

what it means to be able to bring about some state of the world. Rather than
complicate £ORA by introducing yet another primitive modality, I adapt a
well-known definition of ability that was originally proposed by Moore [157].
As a first attempt to define ability, we might say an agent has the ability to
achieve some state <p if it knows of an action that it can perform,which would
be guaranteed to achieve the state of affairs. I will call this first-order ability,
and define it as follows.

((an D i <p) � :30: . (Bel i (Agt 0: i) 1\ (Achvs 0: <p)) 1\ (Agt 0: i) 1\ (Achvs 0: <p)

Notice that the action 0: in this definition is quantified de re with respect to the
Bel modality [101, p.183]. The significance of this is that the agent must be
"aware of the identity" of the action - it must have a rigid designator for it.
Thus it is not enough for the agent to believe that there exists some action that
will achieve the goal. It must be aware of exactly which action will achieve it.
Before proceeding, we prove some results about first-order ability. First,
we show that if an agent has the first-order ability to bring about some state of
affairs,then that state of affairs is actually possible.
Cooperation 15 1

THEOREM 8.1: FS ((an D i <p) => E¢ <p

Proof Assume (M, V,w, t) FS ((an D i <p) for arbitrary (M, V,w, t). By
expanding out the definition of (an D, we get (M, V,w, t) FS 3a· (Achvs a <p).
From this and the definition of Achvs we get (M, V,w, t) FS E (Happens a) /\
A (Happens a; <p?), and hence (M, V,w, t) FS E¢ <p. _

If an agent has first-order ability to bring about a state of affairs,then it is aware

of this.

THEOREM 8.2: FS ((an D i <p) => (Bel i ((an D i <p))

Proof We need to show that if (M, V,w, t) FS ((an D i <p) for arbitrary
(M, V,w,t), then (M, V,w',t) FS ((an D i <p) for all w' E W such that
w' E 8� ([i]). Start by assuming that (M, V,w, t) FS ((an D i <p) for arbitrary
(M, V,w, t). Hence (M, V,w, t) FS 3a · (Bel i (Agt a i) /\ (Achvs a <p)),
and so for all w' E W such that w' E 8� ([i]), we have (M, V,w', t) FS
3a· (Agt a i)/\ (Achvs a <p). We want to show that (M, V,w", t) FS (Agt a i)/\
'
(Achvs a <p) for all w" E W such that w" E 8� ([i]). But since the belief
accessibility relation 8 is transitive, it must be that if w' E 87 ([i]), and w" E
8�' ([i]), then w" E 87 ([i]). Hence (M, V,w", t) FS (Agt a i) /\ (Achvs a <p),
so (M, V,w', t) FS 3a · (Bel i (Agt a i) /\ (Achvs a <p)) /\ (Agt a i) /\
(Achvs a <p), and hence (M, V,w', t) FS ((an D i <p), and we are done. _

If an agent has the first-order ability to bring about some state of affairs, then
it believes this state of affairs is possible.

THEOREM 8.3: FS ((an D i <p) => (Bel i E¢ <p)

Proof Assume that (M, V,w, t) FS ((an D i <p) for arbitrary (M, V,w, t). By
Theorem 8.2, we therefore have (M, V,w', t) FS ((an D i <p) for all w' E W
such that w' E 87 ([i]). From Theorem 8.1, we thus have (M, V,w', t) FS
E¢ <p, and we are done. _

An obvious failing of this definition when measured against our intuitions

about ability is that it does not allow for an agent performing an action in
order to find out how to bring about some state of affairs. To quote a classic
example, I have the ability to telephone my friend Bill even though I do not
know Bill's telephone number; I can perform an action (looking up his number
152 Chapter 8

in the phone book) that will furnish me with the relevant information.
This motivates a definition of k order ability, (k E IN, k > 0), which allows
for the possibility of an agent performing an action in order to find out how to
bring about a state of affairs. The idea is that an agent will have the k order
ability to bring about a state of affairs cp if it has the first-order ability to bring
about k 1 order ability to bring about cpo We formalize this as follows.
-

(Cank i cp) � (Cank-1 i (Can D i cp)) for k > 0

Thus

(Can 1 i cp) � (Can D i (Can D i cp))

(Can 2 i cp) � (Can D i (Can D i (Can D i cp)))
. . . and so on.

We can then define ability simpliciter as k order ability,for some k E IN.

(Can i cp) � V (Cank i cp)

k� D
Notice that Can is essentially an infinitary disjunction; it has the characteristic
of a least fixed point (see the discussion on fixed points in chapter 6). It is easy
to see that first-order ability implies ability simpliciter.

THEOREM 8.4: FS (Can D i cp) => (Can i cp)

We can also prove results analogous to Theorems 8.1, 8.2, and 8.3 for Can.
(Proofs for Theorems 8.5 and 8.6 are straightforward,and are therefore omit­
ted.)

THEOREM 8.5: FS (Can i cp) => EO cp

THEOREM 8.6: FS (Can i cp) => (Bel i EO cp)

THEOREM 8.7: FS (Can i cp) => (Bel i (Can i cp))

Proof The proof is by induction on the degree of ability,k. We show that for
all k E IN, FS (Cank i cp) => (Bel i (Cank i cp)). The inductive base (k 0) =

is simply Theorem 8.2. The inductive assumption is that if (M, V,w, t) FS

(Cank i cp) for arbitrary (M, V,w,t), then (M, V,w',t) FS (Cank i cp) for all
w' E W such that w' E B7 ([i]). The inductive step is then a straightforward
Cooperation 153

generalization of Theorem 8.2. •

To simplify future definitions, we will introduce another derived operator,

Can't, which has the obvious interpretation.

(Can't i 'P) � , (Can i 'P)

We shall assume that if an agent is unable to achieve some state of affairs,then

it is aware that it is unable to achieve this.

FS (Can't i 'P) => (Bel i (Can't i 'P)) (8.1)

We now need to define multiagent ability, which we do by simply adapting the

definition of single-agent ability to the multiagent case.

(J-CanO g'P) � :3a . (M-Belg (Agts ag) 1\ (Achvs a 'P)) 1\

(Agts ag) 1\ (Achvs a 'P)
(J-Cankg'P) � (J-Cank-1 g (J-Cano g'P)) for k > 0
(J-Cang'P) � V k� (J-Cankg'P)
O
We now present some results about joint ability that are analogous to Theo­
rems 8.1 through to 8.7. (The proofs are omitted, since these are straightfor­
ward adaptations of earlier results.)

THEOREM 8. 8: FS (J-Cano g'P) => E <>'P

THEOREM 8.9: FS (J-Cano g'P) => (M-Belg E<>'P)

THEOREM 8.10: FS (J-Cano g'P) => (M-Belg (J-Cano g'P))

THEOREM 8. 11: FS (J-Cano g'P) => (J-Cang'P)

THEOREM 8.12: FS (J-Cang'P) => E <>'P

THEOREM 8. 13: FS (J-Cang'P) => (M-Belg (J-Cang'P))

THEOREM 8. 14: FS (J-Cang'P) => (M-Bel i E<>'P)

154 Chapter 8

Potential for Cooperation Defined

We can now formally define potential for cooperation. With respect to agent
i's desire <p, there is potential for cooperation iff:

1. there is some groupg such that i believes thatg can jointly achieve <p;

and either

2. i can't achieve <p in isolation; or

3. i believes that for every actiona that it could perform that achieves <p, it has
a desire of not performinga.

Note that in clause (1), an agent needs to know the identity of the group that
it believes can cooperate to achieve its desire. This is perhaps an over-strong
assumption. It precludes an agent attempting to find out the identity of a group
that can achieve the desire, and it does not allow an agent to simply broadcast
its desire in the hope of attracting help (as in the Contract Net protocol [222]).
I leave such refinements to future work. Clause (2) represents the paradigm
reason for an agent considering a cooperative solution: because it is unable to
achieve the desire on its own. Clause (3) defines the alternative reason for an
agent considering cooperation: it prefers not to perform any of the actions that
might achieve the desire. (I do not consider the reasons why an agent will not
want to perform a particular action - these will be domain-specific.)
To simplify the presentation that follows slightly, we will make use of an
abbreviation Doesn't, to capture the sense of "something doesn't happen."

(Doesn'ta) � A D-,(Happensa)

Using the various definitions above, we can now formally state the condi­
tions that characterize the potential for cooperation.

(PfC i <p) � (Des i <p) 1\

(Bel i -' <p) 1\
3g· (Bel i (J-Cang <p)) 1\
[(Can'ti<P)V
(Bel iVa· (Agta i) 1\ (Achvsa <p) ::::} (Int i (Doesn'ta))) ]
Cooperation 155

We can prove some properties of potential for cooperation.

THEOREM 8. 15: FS (PfC i 'P) => (Bel i EO'P)

Proof Assume (M, V,w, t) FS (PfC i 'P) for arbitrary (M, V,w, t). By ex­
panding the definition of Pfc, we get (M, V,w, t) FS :3 g . (Bel i (J-Cang'P)).
So for all w' E W such that w' E B7 ([i]), we have (M, V,w', t) FS
'
(J-Can g'P), and so from Theorem 8.12, we have (M, V,w , t) F EO'P. Thus
(M, V,w, t) FS (Bel i EO'P). _

Assuming an agent is aware of its desires,and there is potential for cooperation

with respect to one of its desires, then the agent is aware of this.

THEOREM 8. 16: FS (Pfc i 'P) => (Bel i (Pfc i 'P)) (assuming agent i is
aware of its desires).

Proof Assume (M, V,w, t) FS (Pfc i 'P), for arbitrary (M, V,w, t). We need
to show that (M, V,w, t) FS (Bel i X), for each conjunct in the definition of
potential for cooperation:

• (M, V,w, t) FS (Bel i (Des i 'P))

From the fact that i is aware of its desires.
• (M, V,w, t) FS (Bel i (Bel i -''P))
Immediate from axiom 4 for belief modalities.
• (M, V,w, t) FS (Bel i:3 g . (Bel i (J-Cang'P)))
Immediate from axiom 4 for belief modalities.
• (M, V,w, t) FS (Bel i (Can't i 'P) V (Bel iVa· (Agt a i) 1\ (Achvs a 'P) =>
(Des i (Doesn't a))))
There are two cases to consider. For the first case,assume that (M, V,w, t) FS
(Can't i 'P). Then by the assumption that agents are aware of what they cannot
achieve, (M, V,w, t) FS (Bel i (Can't i 'P)).
For the second case, assume (M, V,w, t) FS (Bel i Va . (Agt a i) 1\
(Achvs a 'P) => (Des i (Doesn't a))). In this case, (M, V,w, t) FS
(Bel i (Bel i Va · (Agt a i) 1\ (Achvs a 'P) => (Des i (Doesn't a)))) fol­
lows from axiom 4 for belief modalities.

This completes the proof. _

156 Chapter 8

Figure 8.2
Agent Al attempts to build a team to help with its problem.

8.3 Team Formation

Having identified the potential for cooperative action with respect to one of its
desires, what is a rational agent to do? Such an agent should attempt to solicit
assistance from a group of agents that it believes can achieve the desire. If the
agent is successful, then at the conclusion of this team formation stage, the
agent will have brought about in such a group a mental state wherein each
member of the group has a nominal commitment to collective action. The
group will not yet have fixed upon an action to perform, and in fact will not
share any kind of commitment other than to the principle of joint action. In
particular, there will not yet be a joint intention: this comes later.
How does an agent go about forming a team? The most important point to
note is that it cannotguarantee that it will be successful in forming a team: it
can only attempt it; see Figure 8.2.
The team formation stage can then be characterized as the following as­
sumption about rational agents: an agent i, who believes that there is potential
for cooperative action with respect to its desire cp, will attempt to bring about
in some groupg (which it believes can jointly achieve cp):
Cooperation 157

1. the mutual intention to achieve rp; and

2. the mutual belief thatg can indeed achieve rp.

In other words,agent i will requestg to carry out rp, and iriformg that they are
able to do it.
It is implicit within this assumption that agents are veracious with respect
to their desires, i.e., that they will try to influence the group by revealing their
true desire. I do not consider cases where agents are mendacious (i.e.,they lie
about their desires), or when agents do not reveal their desires. The interested
reader is referred to [68, pp.159-165] for a discussion and formalization of
such considerations.
It is useful to introduce a definition that captures the commitment that
agents have to collective action if team formation is successful. We write

(PreTeamg rp)

to indicate that: (i) it is mutually believed ing thatg can jointly achieve rp; and
(ii)g mutually intend rp.

(PreTeamg rp) � (M-Belg (J-Cang rp)) 1\ (M-Intg rp)

The following result captures a key property of PreTeam.

THEOREM 8. 17: FS (PreTeamg rp) => (M-Belg EO rp)

Proof Assume that (M, V,w, t) FS (PreTeamgrp) for arbitrary (M, V,w, t).
The definition of PreTeam gives (M, V,w, t) FS (M-Bel g (J-Can g rp)).
Theorem 8.12 tells us that F (J-Can g rp) => EO rp, and from necessitation
for M-Bel operators, we therefore know that FS (M-Bel g ( (J-Can g rp) =>
EO rp)). From the K axiom for M-Bel and propositional reasoning, we can
therefore conclude that (M, V,w, t) FS (M-Belg EO rp). •

We will define a composite speech act FormTeam, the purpose of which is to

bring about a PreTeam mental state.

{FormTeam ig a rp} � {Inform ig a (J-Can g rp)}; {Request ig a rp}

Thus performing an action {FormTeam ig a rp} involves first informingg that

they are able to achieve rp, and then requestingg to achieve rp. (Notice that for
simplicity, I assume the same action a is used for both the inform and request
speech acts.)
158 Chapter 8

Using this composite speech act, we can give the assumption that charac­
terizes the behavior of an agent who recognizes the potential for cooperative
action with respect to one of its desires.

FS Vi· (PfC i cp) => A3g· 3a· (Happens {FormTeam iga cp}) (8.2)

If team formation is successful, then for the first time there will be a social
commitment: a commitment by a group of agents on behalf of another agent.

8.4 Plan Formation

If an agent is successful in its attempt to solicit assistance,then there will be a

group of agents with a mutual intention to cooperative action, and the mutual
belief that they can achieve the desire. But collective action cannot actually
begin until the group agrees on what they will actually do. Hence the next
stage: plan formation.
We saw above that a group will not form a collective unless they believe
they can actually achieve the desire. This, in turn, implies there is at least one
action known to the group that will take them "closer" to the goal (see the
definition of J-Can, above). However,it is possible that there are many agents
that know of actions the group can perform in order to take them closer to the
goal. Moreover, some members of the collective may have objections to one
or more of these actions. One of the desiderata for our model is that agents are
autonomous - they have control over their internal state, and will not simply
perform an action because another agent wants them to [249]. It is therefore
necessary for the collective to come to some agreement about exactly which
course of action they will follow. Such an agreement is reached via negotiation.
Negotiation usually involves agents making reasoned arguments for and
against courses of action; making proposals and counter proposals; suggesting
modifications or amendments to plans; and continuing in this way until all the
participants have agreed on a final result. ! Negotiation has long been recog­
nized as a process of some importance in multiagent systems [228,195,123].
Unfortunately, these analyses demonstrate that negotiation is also extremely
complex - a rigorous attempt at formalization is quite beyond the scope of
this book (see the "Notes and Further Reading" section for references). In­
stead, we simply offer some observations about the weakest conditions under
which negotiation can be said to have occurred.

1 It may also involve agents lying, though we shall not consider such cases here.
Cooperation 159

What can we say about negotiating a plan? First, we note that negotiation
may fail: the collective may simply be unable to reach agreement,due to some
irreconcilable differences.
In this case, the minimum condition required for us to be able to say that
negotiation occurred at all is that at least one agent proposed a course of
action that it believed would take the collective closer to the goal. However,
negotiation may also succeed. In this case, we expect a team action stage to
follow immediately - we shall say no more about team action here, as this is
the subject of the next section.
We shall now make the above discussion more precise. First, we define
joint attempts: what it means for a group of agents to collectively attempt
something. As might be expected,joint attempts are a generalization of single­
agent attempts. An attempt by a group of agentsg to bring about a state 'P is
an action 0:, of whichg are the agents, performed with the mutual desire that
after0: is performed,'P is satisfied,or at least'l/J is satisfied (where'l/J represents
what it takes to make a reasonable effort).

r (M-Belg ""'P) A
(Agts 0:g) A
{J-Attemptgo:'P'l/J}
A

=
(M- D esg (AchvSO:'P)) A
(M-Intg (Achvs 0:;'l/J?))

We can now state the minimum conditions required for negotiation to have
occurred. Intuitively,the group will try to bring about a state where they have
agreed on a common plan,and intend to act on it. Failing that, they will bring
about a state where at least one of them has proposed a plan that it believed
would achieve the desired goal. More formally,if groupg are a pre-team with
respect to agent i's desire 'P, then g will eventually jointly attempt to bring
about a state in whichg are a team with respect to i's desire 'P, or,failing that,
to at least bring about a state where some agent j E g has made g mutually
aware of its belief that some action0: can be performed byg in order to achieve
'P. Formally,this assumption is as follows:

FS (PreTeamg'P) => A<>30: · (Happens {J-Attemptg0:'l/J X}) (8.3)

where

'l/J � (Teamsocg'P'l/Jsoc)
160 Chapter 8

and

x � 3j· 3a . (j E g) 1\ (M-Belg (Belj (Agtsag) 1\ (Achvsa <p))).

Recall that Teamsoc was defined in chapter 6,to capture the notion of a "team"
mental state.
We can make some other assumptions about agent behavior during nego­
tiation. Most importantly, we assume that agents will attempt to bring about
their preferences. For example,if an agent has an objection to some plan,then
it will attempt to prevent this plan being carried out. Similarly,if it has a prefer­
ence for some plan,then it will attempt to bring this plan about. More precisely,
if groupg are a pre-team with respect to agent i's desire <p, and there is some
actiona such that it is mutually believed ing thata achieves <p, and thatg are
the agents ofa, then every agentj E g that has a preference thata does/does
not occur will attempt to ensure thata does/does not occur,by at least making
g mutually aware of its preference for/againsta. Note that we are once again
assuming that agents are veracious; that they attempt to influence the team by
revealing their true preferences, rather than by lying about their preferences,
or not revealing their true preferences. An agent can attempt to cause a team
to perform some action by simply making use of the RequestTh speech act,as
defined in chapter 7.
If the plan formation phase is successful then the team will have a full joint
commitment to the joint goal,and will have agreed to the means by which they
will pursue their joint goal. Ideally,we would also like to specify that the group
negotiate a convention for monitoring progress of the team action, but this is
beyond the scope of this book.
If a collective is successful in its attempt to negotiate a plan,then we expect
that collective to follow up negotiation with action. This gives us the fourth,
and final stage in our model: team action. Team action will take place when
the group of agents have a joint intention towards some state of affairs, have
a mutually agreed action that will achieve this state of affairs, and have some
convention in place to monitor the progress of the team action. We have already
seen examples of this type of team action in chapter 6.

8.5 Notes and Further Reading

Many attempts to formally define ability have appeared in the literature. The
earliest discussion on the subject in the artificial intelligence literature is prob-
Cooperation 16 1

ably that by McCarthy and Hayes, in their well-known work [151]. The best­
known definition of ability,(and the one upon which the definition in this book
is based) is that developed by Robert Moore [156, 157]. Other examples in­
clude the work of Brown [25], Singh's definitions of ability and cooperative
ability [212,216],and the work of Werner [234] and Morgenstern [160,161].
The model of cooperative problem solving presented here is based on
one I developed jointly with Nick Jennings [247, 248, 250]. These papers
were the first to present the four-stage model of cooperative problem solving,
and the first to formally define such notions as potential for cooperation.
The logic used in this earlier work is an ancestor of .cORA, with the main
difference being that .cORA uses the BOI model of Rao and Georgeff; in
the original formulation, the logic more closely resembled that of Cohen and
Levesque [35,36].
Afsaneh Haddadi, building on [247,248,250], developed a formalization
of cooperation plans [86,87]. She further developed the concept of potential
for cooperation that was presented in [247, 248, 250], and discussed how
cooperation plans specified in a BOI logic might be implemented in a practical
BOI system.
A number of attempts have appeared in the literature to formalize multi­
agent argumentation. One of the first was [124]; since then, Parsons and col­
leagues have presented several such models [169,207]. Kraus et al. present a
detailed logical formalization of argumentation using a BOI-like logic in [125].
The best-developed alternative to the model of cooperative action pre­
sented in this chapter is the "SharedPlans" model developed by Grosz and
Kraus [84, 85]. Expressed in a rich logic that combines multimodal and first­
order elements, the primary emphasis in SharedPlans is on giving an account
of how the plans of individual agents can be reconciled with those of a team.
9 Logic and Agent Theory

Mathematical logic is one of the most important tools available in computer

science and AI today. In this chapter, we will step back from our attempts to
formalize the properties of rational agents to see what role logic can play in
the engineering of rational agents.
Broadly speaking, logic has played a role in three aspects of software
engineering:

• as a specification language;
• as a programming language; and
• as a verification language.

In the sections that follow, we will discuss the use of logics such as CORA in
these three processes.

9.1 Specification

The software development process begins by establishing the client's require­

ments. When this process is complete, a specification is developed, which sets
out the functionality of the new system. Temporal and dynamic logics, closely
related to CORA, have found wide applicability in the specification of sys­
tems. An obvious question is therefore whether logics such as CORA might
be used as specification languages.
A specification expressed in CORA would be a formula 'P. The idea is
that such a specification would express the desirable behavior of a system. To
see how this might work, consider the following formula of CORA, intended
to form part of a specification of a process control system.

(Bel i Open(valve32)) => (Int i (Bel) Open(valve32)))

This formula says that if i believes valve 32 is open, then i should intend that)
believes valve 32 is open. A rational agent i with such an intention can select
a speech act to perform in order to inform) of this state of affairs. It should be
intuitively clear how a system specification might be constructed using such
formulae, to define the intended behavior of a system.
One of the main desirable features of a software specification language
is that it should not dictate how a specification should be satisfied by an
implementation. It should be clear that the specification above has exactly these
164 Chapter 9

properties. It does not dictate how agent i should go about makingj aware that
valve 32 is open. We simply expect i to behave as a rational agent given such
an intention.
There are a number of problems with the use of languages such as CORA
for specification. The most worrying of these is with respect to their semantics.
The semantics for the modal connectives (for beliefs, desires, and intentions)
are given in the normal modal logic tradition of possible worlds [31]. So, for
example, an agent's beliefs in some state are characterized by a set of different
states, each of which represents one possibility for how the world could actu­
ally be, given the information available to the agent. In much the same way,
an agent's desires in some state are characterized by a set of states that are
consistent with the agent's desires. Intentions are represented similarly. There
are several advantages to the possible worlds model: it is well studied and well
understood, and the associated mathematics of correspondence theory is ex­
tremely elegant. These attractive features make possible worlds the semantics
of choice for almost every researcher in formal agent theory. However, there
are also a number of serious drawbacks to possible worlds semantics. First,
possible worlds semantics imply that agents are logically perfect reasoners, (in
that their deductive capabilities are sound and complete), and they have infi­
nite resources available for reasoning. No real agent, artificial or otherwise, has
these properties. (See appendix B for a discussion of this.)
Second, possible worlds semantics are generally ungrounded. That is, there
is usually no precise relationship between the abstract accessibility relations
that are used to characterize an agent's state, and any concrete computational
model. As we shall see in later sections, this makes it difficult to go from
a formal specification of a system in terms of beliefs, desires, and so on, to
a concrete computational system. Similarly, given a concrete computational
system, there is generally no way to determine what the beliefs, desires, and
intentions of that system are. If temporal modal logics such as CORA are to
be taken seriously as specification languages, then this is a significant problem.

9.2 Implementation

Specification is not (usually!) the end of the story in software development.

Once given a specification, we must implement a system that is correct with
respect to this specification. The next issue we consider is this move from
abstract specification to concrete computational model. There are at least three
Logic and Agent Theory 165

possibilities for achieving this transformation:

1. manually refine the specification into an executable form via some princi­
pled but informal refinement process (as is the norm in most current software
development);
2. directly execute or animate the abstract specification; or
3. translate or compile the specification into a concrete computational form
using an automatic translation technique.

In the subsections that follow, we shall investigate each of these possibilities

in turn .

Refinement

At the time of writing, most software developers use structured but informal
techniques to transform specifications into concrete implementations. Probably
the most common techniques in widespread use are based on the idea of top­
down refinement. In this approach, an abstract system specification is refined
into a number of smaller, less abstract subsystem specifications, which together
satisfy the original specification. If these subsystems are still too abstract to be
implemented directly, then they are also refined. The process recurses until the
derived subsystems are simple enough to be directly implemented. Through­
out, we are obliged to demonstrate that each step represents a true refinement of
the more abstract specification that preceded it. This demonstration may take
the form of a formal proof, if our specification is presented in, say, Z [224]
or VDM [110]. More usually, justification is by informal argument. Object­
oriented analysis and design techniques, which also tend to be structured but
informal, are also increasingly playing a role in the development of systems
(see, e.g., [19]).
For functional systems, which simply compute a function of some input
and then terminate, the refinement process is well understood, and compar­
atively straightforward. Such systems can be specified in terms of pre- and
post-conditions (e.g., using Hoare logic [97]). Refinement calculi exist, which
enable the system developer to take a pre- and post-condition specification,
and from it systematically derive an implementation through the use of proof
rules [159]. Part of the reason for this comparative simplicity is that there is of­
ten an easily understandable relationship between the pre- and post-conditions
that characterize an operation and the program structures required to imple­
ment it.
166 Chapter 9

For agent systems, which fall into the category of Pnuelian reactive sys­
tems (see the discussion in chapter 1), refinement is not so straightforward.
This is because such systems must be specified in terms of their ongoing be­
havior - they cannot be specified simply in terms of pre- and post-conditions.
In contrast to pre- and post-condition formalisms, it is not so easy to determine
what program structures are required to realize such specifications. As a conse­
quence, researchers have only just begun to investigate refinement and design
technique for agent-based systems.

Case Study: A Methodology for BDI Agents

In [117], Kinny et al. propose a four-stage design methodology for systems of

8DI agents. The methodology is closely linked to a specific realization of the
8DI model: the PRS architecture [78]. The methodology may be summarized
as follows:

1. Identify the relevant roles in the application domain, and on the basis of
these, develop an agent class hierarchy. An example role might be weather
monitor, whereby agent i is required to make agent j aware of the prevailing
weather conditions every hour.
2. Identify the responsibilities associated with each role, the services required
by and provided by the role, and then determine the goals associated with each
service. With respect to the above example, the goals would be to find out the
current weather, and to make agentj aware of this information.
3. For each goal, determine the plans that may be used to achieve it, and the
context conditions under which each plan is appropriate. With respect to the
above example, a plan for the goal of making agent j aware of the weather
conditions might involve sending a message toj.
4. Determine the belief structure of the system - the information require­
ments for each plan and goal. With respect to the above example, we might
propose a unary predicate WindSpeed(x) to represent the fact that the current
wind speed is x. A plan to determine the current weather conditions would
need to be able to represent this information.

Note that the analysis process will be iterative, as in more traditional method­
ologies. The outcome will be a model that closely corresponds to the PRS agent
architecture. As a result, the move from end-design to implementation using
PRS is relatively simple.
Logic and Agent Theory 167

Directly Executing Agent Specifications

One major disadvantage with manual refinement methods is that they intro­
duce the possibility of error. If no proofs are provided, to demonstrate that
each refinement step is indeed a true refinement, then the correctness of the
implementation process depends upon little more than the intuitions of the de­
veloper. This is clearly an undesirable state of affairs for applications in which
correctness is a major issue. One possible way of circumventing this problem,
which has been widely investigated in mainstream computer science, is to get
rid of the refinement process altogether, and directly execute the specification.
It might seem that suggesting the direct execution of complex agent speci­
fication languages is naive - it is exactly the kind of suggestion that detractors
of logic-based AI hate. One should therefore be very careful about what claims
or proposals one makes. However, in certain circumstances, the direct execu­
tion of agent specification languages is possible.
What does it mean, to execute a formula <p of logic L? It means generating
a logical model, M, for <p, such that M F <p [64]. If this could be done without
interference from the environment - if the agent had complete control over its
environment - then execution would reduce to constructive theorem-proving,
where we show that <p is satisfiable by building a model for <po In reality, of
course, agents are not interference-free: they must iteratively construct a model
in the presence of input from the environment. Execution can then be seen as
a two-way iterative process:

• environment makes something true;

• agent responds by doing something, i.e., making something else true in the

model;
• environment responds, making something else true;
•

Execution of logical languages and theorem-proving are thus closely related.

This tells us that the execution of sufficiently rich (quantified) languages is not
possible (since any language equal in expressive power to first-order logic is
undecidable).
A useful way to think about execution is as if the agent is playing a game
against the environment. The specification represents the goal of the game: the
agent must keep the goal satisfied, while the environment tries to prevent the
agent from doing so. The game is played by agent and environment taking
turns to build a little more of the model. If the specification ever becomes
168 Chapter 9

false in the (partial) model, then the agent loses. In real reactive systems,
the game is never over: the agent must continue to play forever. Of course,
some specifications (logically inconsistent ones) cannot ever be satisfied. A
winning strategy for building models from (satisfiable) agent specifications in
the presence of arbitrary input from the environment is an execution algorithm
for the logic.

Case Study: Concurrent METATEM

Concurrent METATEM is a programming language for multiagent systems,

that is based on the idea of directly executing linear time temporal logic agent
specifications [65, 63]. A Concurrent METATEM system contains a number
of concurrently executing agents, each of which is programmed by giving it
a temporal logic specification of the behavior it is intended the agent should
exhibit. An agent specification has the form Ai Pi => Fi, where Pi is a temporal
logic formula referring only to the present or past, and Fi is a temporal logic
formula referring to the present or future. The Pi => Fi formulae are known as
rules. The basic idea for executing such a specification may be summed up in
the following slogan:

on the basis of the past do the future.

Thus each rule is continually matched against an internal, recorded history,

and if a match is found, then the rule fires. If a rule fires, then any variables
in the future time part are instantiated, and the future time part then becomes
a commitment that the agent will subsequently attempt to satisfy. Satisfying
a commitment typically means making some predicate true within the agent.
Here is a simple example of a Concurrent METATEM agent definition:

Oask(x) => Ogive(x)

(...,ask(x) Z (give(x) 1\ ...,ask(x))) => ...,give(x)
give(x) 1\ give(y) => (x y)=

The agent in this example is a controller for a resource that is infinitely

renewable, but which may only be possessed by one agent at any given time.
The controller must therefore enforce mutual exclusion. The predicate ask(x)
means that agent x has asked for the resource. The predicate give(x) means
that the resource controller has given the resource to agent x. The resource
controller is assumed to be the only agent able to "give" the resource. However,
many agents may ask for the resource simultaneously. The three rules that
define this agent's behavior may be summarized as follows:
Logic and Agent Theory 169

• Rule 1: if someone asks, then eventually give;

• Rule 2: don't give unless someone has asked since you last gave; and
• Rule 3: if you give to two people, then they must be the same person (i.e.,

don't give to more than one person at a time).

Concurrent METATEM agents can communicate by asynchronous broadcast

message passing, though the details are not important here.

Compiling Agent Specifications

An alternative to direct execution is compilation. In this scheme, we take our

abstract specification, and transform it into a concrete computational model
via some automatic synthesis process. The main perceived advantages of com­
pilation over direct execution are in run-time efficiency. Direct execution of
an agent specification, as in Concurrent METATEM, above, typically involves
manipulating a symbolic representation of the specification at run time. This
manipulation generally corresponds to reasoning of some form, which is com­
putationally costly (and in many cases, simply impracticable for systems that
must operate in anything like real time). In contrast, compilation approaches
aim to reduce abstract symbolic specifications to a much simpler computa­
tional model, which requires no symbolic representation. The "reasoning"
work is thus done off-line, at compile-time; execution of the compiled system
can then be done with little or no run-time symbolic reasoning. As a result,
execution is much faster. The advantages of compilation over direct execution
are thus those of compilation over interpretation in mainstream programming.
Compilation approaches usually depend upon the close relationship be­
tween models for temporallmodal logic (which are typically labeled graphs of
some kind), and automata-like finite state machines. Crudely, the idea is to take
a specification <p, and do a constructive proof of the implementability of <p,
wherein we show that the specification is satisfiable by systematically attempt­
ing to build a model for it. If the construction process succeeds, then the spec­
ification is satisfiable, and we have a model to prove it. Otherwise, the speci­
fication is unsatisfiable. If we have a model, then we "read off" the automaton
that implements <p from its corresponding model. The most common approach
to constructive proof is the semantic tableaux method of Smullyan [223].
In mainstream computer science, the compilation approach to automatic
program synthesis has been investigated by a number of researchers. Perhaps
the closest to our view is the work of Pnueli and Rosner [178] on the automatic
synthesis of reactive systems from branching time temporal logic specifica-
170 Chapter 9

tions. The goal of their work is to generate reactive systems, which share many
of the properties of our agents (the main difference being that reactive systems
are not generally required to be capable of rational decision making in the way
we described above). To do this, they specify a reactive system in terms of a
first-order branching time temporal logic formula \:Ix 3y Acp ( x,y) : the predicate
cp characterizes the relationship between inputs to the system (x) and outputs
(y). Inputs may be thought of as sequences of environment states, and outputs
as corresponding sequences of actions. The A is the universal path quantifier.
The specification is intended to express the fact that in all possible futures, the
desired relationship cp holds between the inputs to the system, x, and its out­
puts, y. The synthesis process itself is rather complex: it involves generating a
Rabin tree automaton, and then checking this automaton for emptiness. Pnueli
and Rosner show that the time complexity of the synthesis process is double
exponential in the size of the specification, i.e., 0(22cOn), where c is a constant
and n = Icpl is the size of the specification cpo The size of the synthesized
program (the number of states it contains) is of the same complexity.
The Pnueli-Rosner technique is rather similar to (and in fact depends upon)
techniques developed by Wolper, Vardi, and colleagues for synthesizing Biichi
automata from linear temporal logic specifications [232]. Biichi automata are
those that can recognize w-regular expressions: regular expressions that may
contain infinite repetition. A standard result in temporal logic theory is that a
formula cp of linear time temporal logic is satisfiable if and only if there exists
a Biichi automaton that accepts just the sequences that satisfy cpo Intuitively,
this is because the sequences over which linear time temporal logic is inter­
preted can be viewed as w-regular expressions. This result yields a decision
procedure for linear time temporal logic: to determine whether a formula cp
is satisfiable, construct an automaton that accepts just the (infinite) sequences
that correspond to models of cp; if the set of such sequences is empty, then cp
is unsatisfiable. The technique for constructing an automaton from the corre­
sponding formula is closely based on Wolper's tableau proof method for tem­
poral logic [236].
Similar automatic synthesis techniques have also been deployed to develop
concurrent system skeletons from temporal logic specifications. Manna and
Wolper present an algorithm that takes as input a linear time temporal logic
specification of the synchronization part of a concurrent system, and generates
as output a program skeleton (based upon Hoare's csp formalism [98]) that re­
alizes the specification [146]. The idea is that the functionality of a concurrent
system can generally be divided into two parts: a functional part, which actu-
Logic and Agent Theory 17 1

ally performs the required computation in the program, and a synchronization

part, which ensures that the system components cooperate in the correct way.
For example, the synchronization part will be responsible for any mutual ex­
clusion that is required. The synthesis algorithm (like the synthesis algorithm
for Biichi automata, above) is based on Wolper's tableau proof method for tem­
poral logic [236]. Very similar work is reported by Clarke and Emerson [32]:
they synthesize synchronization skeletons from branching time temporal logic
(CTL) specifications.

Case Study: Situated Automata

Perhaps the best-known example of this approach to agent development is the

situated automata paradigm of Rosenschein and Kaelbling [196, 114, 197]. In
this approach, an agent has two main components:

• a perception part, which is responsible for observing the environment and

updating the internal state of the agent; and

• an action part, which is responsible for deciding what action to perform,

based on the internal state of the agent.

Rosenschein and Kaelbling developed two programs to support the develop­

ment of the perception and action components of an agent respectively. The
RULER program takes a declarative perception specification and compiles it
down to a finite state machine. The specification is given in terms of a theory
of knowledge. The semantics of knowledge in the declarative specification lan­
guage are given in terms of possible worlds, in the way described above. Cru­
cially, however, the possible worlds underlying this logic are given a precise
computational interpretation, in terms of the states of a finite state machine. It
is this precise relationship that permits the synthesis process to take place.
The action part of an agent in Rosenschein and Kaelbling's framework is
specified in terms of goal reduction rules, which encode information about
how to achieve goals. The GAPPS program takes as input a goal specification,
and a set of goal reduction rules, and generates as output a set of situation
action rules, which may be thought of as a lookup table, defining what the
agent should do under various circumstances, in order to achieve the goal. The
process of deciding what to do is then very simple in computational terms,
involving no reasoning at all. (A similar technique, called universal plans, was
developed by Schoppers [201].)
172 Chapter 9

9.3 Verification

Once we have developed a concrete system, we need to show that this sys­
tem is correct with respect to our original specification. This process is known
as verification, and it is particularly important if we have introduced any in­
formality into the development process. For example, any manual refinement,
done without a formal proof of refinement correctness, creates the possibility
of a faulty transformation from specification to implementation. Verification
is the process of convincing ourselves that the transformation was sound. We
can divide approaches to the verification of systems into two broad classes:
(1) axiomatic; and (2) semantic (model checking). In the subsections that fol­
low, we shall look at the way in which these two approaches have evidenced
themselves in agent-based systems.

Axiomatic Approaches

Axiomatic approaches to program verification were the first to enter the main­
stream of computer science, with the work of Hoare in the late 1960s [97].
Axiomatic verification requires that we can take our concrete program, and
from this program systematically derive a logical theory that represents the be­
havior of the program. Call this the program theory. If the program theory is
expressed in the same logical language as the original specification, then veri­
fication reduces to a proof problem: show that the specification is a theorem of
(equivalently, is a logical consequence of) the program theory.
The development of a program theory is made feasible by axiomatizing
the programming language in which the system is implemented. For example,
Hoare logic gives us more or less an axiom for every statement type in a simple
PASCAL-like language. Once given the axiomatization, the program theory can
be derived from the program text in a systematic way.
Perhaps the most relevant work from mainstream computer science is the
specification and verification of reactive systems using temporal logic, in the
way pioneered by Pnueli, Manna, and colleagues [145]. The idea is that the
computations of reactive systems are infinite sequences, which correspond
to models for linear temporal logic.! Temporal logic can be used both to
develop a system specification, and to axiomatize a programming language.
This axiomatization can then be used to systematically derive the theory of a

1 The set of all computations of a reactive system is a tree-like structure, corresponding to a model
for branching time temporal logic [50].
Logic and Agent Theory 173

program from the program text. Both the specification and the program theory
will then be encoded in temporal logic, and verification hence becomes a proof
problem in temporal logic.
Comparatively little work has been carried out within the agent-based
systems community on axiomatizing multiagent environments. I shall review
just one approach.

Case Study: Axiomatizing Two Multiagent Languages

In [237], an axiomatic approach to the verification of multiagent systems

was proposed. Essentially, the idea was to use a temporal belief logic to
axiomatize the properties of two multiagent programming languages. Given
such an axiomatization, a program theory representing the properties of the
system could be systematically derived in the way indicated above.
A temporal belief logic was used for two reasons. First, a temporal com­
ponent was required because, as we observed above, we need to capture the
ongoing behavior of a multiagent system. A belief component was used be­
cause the agents we wish to verify are each symbolic AI systems in their own
right. That is, each agent is a symbolic reasoning system, which includes a
representation of its environment and desired behavior. A belief component in
the logic allows us to capture the symbolic representations present within each
agent.
The two multiagent programming languages that were axiomatized in the
temporal belief logic were Shoham's AGENTO [206], and Fisher's Concurrent
METATEM (see above). The basic approach was as follows:

1. First, a simple abstract model was developed of symbolic AI agents. This

model captures the fact that agents are symbolic reasoning systems, capable
of communication. The model gives an account of how agents might change
state, and what a computation of such a system might look like.
2. The histories traced out in the execution of such a system were used as
the semantic basis for a temporal belief logic. This logic allows us to express
properties of agents modeled at stage (1).
3. The temporal belief logic was used to axiomatize the properties of a mul­
tiagent programming language. This axiomatization was then used to develop
the program theory of a multiagent system.
4. The proof theory of the temporal belief logic was used to verify properties
of the system [246].
174 Chapter 9

Note that this approach relies on the operation of agents being sufficiently
simple that their properties can be axiomatized in the logic. It works for
Shoham's AGENTO and Fisher's Concurrent METATEM largely because these
languages have a simple semantics, closely related to rule-based systems,
which in turn have a simple logical semantics. For more complex agents,
an axiomatization is not so straightforward. Also, capturing the semantics of
concurrent execution of agents is not easy (it is, of course, an area of ongoing
research in computer science generally).

Semantic Approaches: Model Checking

Ultimately, axiomatic verification reduces to a proof problem. Axiomatic ap­

proaches to verification are thus inherently limited by the difficulty of this
proof problem. Proofs are hard enough, even in classical logic; the addition
of temporal and modal connectives to a logic makes the problem considerably
harder. For this reason, more efficient approaches to verification have been
sought. One particularly successful approach is that of model checking [33].
As the name suggests, whereas axiomatic approaches generally rely on syn­
tactic proof, model-checking approaches are based on the semantics of the
specification language.
The model-checking problem, in abstract, is quite simple: given a formula
'P of language L, and a model M for L, determine whether or not 'P is valid
in M, i.e., whether or not M FL 'P. Verification by model checking has been
studied in connection with temporal logic [134]. The technique once again
relies upon the close relationship between models for temporal logic and finite­
state machines. Suppose that 'P is the specification for some system, and 1f is a
program that claims to implement 'P. Then, to determine whether or not 1f truly
implements 'P, we proceed as follows:

• take 1f, and from it generate a model M7r that corresponds to 1f, in the sense

that M7r encodes all the possible computations of 1f;

• determine whether or not M7r F 'P, i.e., whether the specification formula
'P is valid in M7r; the program 1f satisfies the specification 'P just in case the
answer is "yes."

The main advantage of model checking over axiomatic verification is in com­

plexity: model checking using the branching time temporal logic eTL [32] can
be done in time O(I'PI x IMI), where I'PI is the size of the formula to be checked,
and IMI is the size of the model against which 'P is to be checked - the number
Logic and Agent Theory 175

of states it contains.

Case Study: Model Checking BDI Systems

In [189], Rao and Georgeff present an algorithm for model checking 8DI sys­
tems. More precisely, they give an algorithm for taking a logical model for
their (propositional) 8DI logic, and a formula of the language, and determin­
ing whether the formula is valid in the model. The technique is closely based
on model-checking algorithms for normal modal logics [92]. They show that
despite the inclusion of three extra modalities (for beliefs, desires, and inten­
tions) into the CTL branching time framework, the algorithm is still quite effi­
cient, running in polynomial time. So the second step of the two-stage model­
checking process described above can still be done efficiently. Similar algo­
rithms have been reported for 8DI-like logics in [13].
The main problem with model-checking approaches for 8DI is that it is
not clear how the first step might be realized for 8DI logics. Where does the
logical model characterizing an agent actually come from? Can it be derived
from an arbitrary program 7r, as in mainstream computer science? To do this,
we would need to take a program implemented in, say, PASCAL, and from it
derive the belief-, desire-, and intention-accessibility relations that are used
to give a semantics to the 8DI component of the logic. Because, as we noted
earlier, there is no clear relationship between the 8DI logic and the concrete
computational models used to implement agents, it is not clear how such a
model could be derived.

9.4 Notes and Further Reading

This chapter is an updated and extended version of [240], which examined

the possibility of using logic to engineer agent-based systems. Since this
article was published, several other authors have proposed the use of agents
in software engineering (see, e.g., [107]).
Structured but informal refinement techniques are the mainstay of real­
world software engineering. If agent-oriented techniques are ever to become
widely used outside the academic community, then informal, structured meth­
ods for agent-based development will be essential. One possibility for such
techniques, followed by Luck and d'Invemo, is to use a standard specifica­
tion technique (in their case, z), and use traditional refinement methods (in
their case, object-oriented development) to transform the specification into an
176 Chapter 9

implementation [141]. This approach has the advantage of being familiar to

a much larger user-base than entirely new techniques, but suffers from the
disadvantage of presenting the user with no features that make it particularly
well-suited to agent specification. It seems certain that there will be much more
work on manual refinement techniques for agent-based systems in the imme­
diate future, but exactly what form these techniques will take is not clear.
With respect to the possibility of directly executing agent specifications, a
number of problems suggest themselves. The first is that of finding a concrete
computational interpretation for the agent specification language in question.
To see what we mean by this, consider models for the agent specification lan­
guage in Concurrent METATEM. These are very simple: essentially just linear
discrete sequences of states. Temporal logic is (among other things) simply a
language for expressing constraints that must hold between successive states.
Execution in Concurrent METATEM is thus a process of generating constraints
as past-time antecedents are satisfied, and then trying to build a next state that
satisfies these constraints. Constraints are expressed in temporal logic, which
implies that they may only be in certain, regular forms. Because of this, it is
possible to devise an algorithm that is guaranteed to build a next state if it is
possible to do so. Such an algorithm is described in [8].
The agent specification language upon which Concurrent METATEM is
based thus has a concrete computational model, and a comparatively simple
execution algorithm. Contrast this state of affairs with languages like CORA,
where we have not only a temporal dimension to the logic, but also modalities
for referring to beliefs, desires, and so on. In general, models for these logics
have ungrounded semantics. That is, the semantic structures that underpin
these logics (typically accessibility relations for each of the modal operators)
have no concrete computational interpretation. As a result, it is not clear how
such agent specification languages might be executed.
Another obvious problem is that execution techniques based on theorem­
proving are inherently limited when applied to sufficiently expressive (first­
order) languages, as first-order logic is undecidable. However, complexity is a
problem even in the propositional case. For "vanilla" propositional logic, the
decision problem for satisfiability is NP-complete [54, p.72]; for linear tempo­
ral logic, the problem is PS PAcE-complete [220]; for simple modal logics of
knowledge, the problem is NP-complete, and for more complex modal logics
of knowledge, the problem is EXPTIME-complete [54, p.73]; for logics that
combine temporal and (S5) modal aspects, the decision problem varies from
PS PAcE-complete in the simplest case to lI t -complete, (and hence undecid-
Logic and Agent Theory 177

able) in the propositional case, depending on what semantic assumptions are

made [54, p.289].
Despite these problems, the undoubted attractions of direct execution have
led to a number of attempts to devise executable logic-based agent languages.
Rao, one of the main figures behind BDl, proposed an executable subset of BDl
logic in his AGENTS PEAK (L ) language [183]. Building on this work, Hindriks
and colleagues developed the 3APL agent programming language [94, 95].
Lesperance, Reiter, Levesque, and colleagues developed the GOLOG language
throughout the latter half of the 1990s as an executable subset of the situa­
tion calculus [130, 131]. Fagin and colleagues have proposed knowledge-based
programs as a paradigm for executing logical formulae which contain epis­
temic modalities [54, 55]. Although considerable work has been carried out
on the properties of knowledge-based programs, comparatively little research
to date has addressed the problem of how such programs might be actually
executed.
Turning to automatic synthesis, we find that the techniques described
above have been developed primarily for propositional specification languages.
If we attempt to extend these techniques to more expressive, first-order spec­
ification languages, then we again find ourselves coming up against the un­
decidability of quantified logic. Even in the propositional case, the theoretical
complexity of theorem-proving for modal and temporal logics is likely to limit
the effectiveness of compilation techniques: given an agent specification of size
1,000, a synthesis algorithm that runs in exponential time when used off-line
is no more useful than an execution algorithm that runs in exponential time
on-line.
Another problem with respect to synthesis techniques is that they typically
result in finite-state, automata-like machines, which are less powerful than
Turing machines. In particular, the systems generated by the processes outlined
above cannot modify their behavior at run-time. In short, they cannot learn.
While for many applications, this is acceptable - even desirable - for equally
many others, it is not. In expert assistant agents, of the type described in [142],
learning is pretty much the raison d'etre. Attempts to address this issue are
described in [113].
Turning to verification, axiomatic approaches suffer from two main prob­
lems. First, the temporal verification of reactive systems relies upon a simple
model of concurrency, where the actions that programs perform are assumed
to be atomic. We cannot make this assumption when we move from programs
to agents. The actions we think of agents as performing will generally be much
178 Chapter 9

more coarse-grained. As a result, we need a more realistic model of concur­

rency. One possibility, investigated in [238], is to model agent execution cy­
cles as intervals over the real numbers, in the style of the temporal logic of
reals [10]. The second problem is the difficulty of the proof problem for agent
specification languages. The theoretical complexity of proof for many of these
logics is quite daunting.
Hindriks and colleagues have used Plotkin's structured operational seman­
tics to axiomatize their 3APL language [94, 95].
With respect to model-checking approaches, the main problem, as we indi­
cated above, is again the issue of ungrounded semantics for agent specification
languages. If we cannot take an arbitrary program and say, for this program,
what its beliefs, desires, and intentions are, then it is not clear how we might
verify that this program satisfied a specification expressed in terms of such
constructs.
A
Appendix:
Summary of Notation

The fonnal tools used in this book are loosely based on the YDM specification
language: the first seven chapters of Jones [110] cover all the necessary ma­
terial. However, anyone familiar with basic set notation and logic should have
no difficulty.
A central idea is that of the type: for the purposes of this book, a type can
be thought of as a (possibly infinite) set. Two "standard " types are assumed:

IN = {a,1,2, ...} the natural numbers

IR the real numbers

The set of subsets of a type T (Le., the powerset of T) is given by p( T). Thus

p({1,2,3}) = {0,{1},{2},{3},{1,2},{1,3},{2,3},{1,2,3}}
The set T of all ordered n-tuples whose first element is of type T1, second
element is of type T2, ... , dh element is of type Tn, may be defined in two
ways. The first is to define it as the cross product of its component types:

T = Tl X ... x Tn

The alternative is to name the elements:

where tl E T1, t2 E T2, ..., tn E Tn. In general, the latter method is used for
large or complex types.
Functions are specified by giving their signature:

f : Dl x . . . x Dn ---+ R

This says that the function f takes n arguments, the first of type (or from
domain) D1, . . . , the dh from domain Dn, and returns a value of type (orfrom
the range) R.
The symbol

should be read "is defined as. " An expression of the fonn

f(x) � e(x)

means that the left hand sidef(x) is defined as the right hand side e(x). For
180 Appendix A

example, consider the definition of material implication:

This definition tells us that a formula with the structure 'P ::::} 'l/J should be
understood as an abbreviation for ''P V 'l/J. It may be useful to think of such
definitions as behaving like macro definitions in a programming language like
c.
A useful operator on functions is to overwrite them. To do this, we use the
dagger operator, "f' Thus by the expressionft {x I-t y} we mean the function
that is the same asf except that it maps x to y. Thus if

f(2) = 3
and

g�ft {2 I-t 7}
then g(2) = 7.
Two standard operations on functions and relations are assumed: dom
takes a function or relation and returns the set of elements in its domain for
which it is defined; ran takes a function or relation and returns the set of
elements in its range for which the corresponding elements in the domain are
defined.
A sequence can be thought of simply as a list of objects, each of which is
of the same type. We denote the set of (possibly empty) sequences over a set T
by T*.
Finally, in this book, we deal with several languages: the symbols 'P, 'l/J,
and X are used as meta-variables, ranging over formulae of these languages.
B
Appendix:
Formal Foundations

In this appendix, I present an overview of the main logical tools available

for reasoning about time and mental states. I assume the reader has some
familiarity with propositional and first-order logic, as well as basic set notation
(see appendix a).

B.t Reasoning about Intentional Notions

Suppose one wishes to reason about belief in a logical framework. Consider

the following simple statement about belief (adapted from [73 , pp.210-211]):

Janine believes Cronos is the father of Zeus. (B.1)

A naive attempt to translate (B.1) into first-order logic might result in the
following:

Bel(Janine, Father (Zeus, Cronos) ) (B.2)

Unfortunately, this naive translation does not work, for at least two reasons.
The first is syntactic: the second argument to the Bel predicate is aformula of
first-order logic, and is not, therefore, a tenn. So (B.2) is not a well-fonned
fonnula of classical first-order logic. The second problem is semantic. The
constants Zeus and Jupiter, by any reasonable interpretation, denote the same
individual: the supreme deity of the classical world. It is therefore acceptable
to write, in first-order logic:

(Zeus = Jupiter) . (B.3 )

Given (B.2) and (B.3 ), the standard rules of first-order logic would allow the
derivation of the following:

Bel(Janine, Father (Jupiter , Cronos) ) (B.4)

But intuition rejects this derivation as invalid: believing that the father of Zeus
is Cronos is not the same as believing that the father of Jupiter is Cronos.
So what is the problem? Why does first-order logic fail here? The problem
is that the intentional notions - such as belief and desire - are referentially
opaque, in that they set up opaque contexts, in which the standard substitution
rules of first-order logic do not apply. In classical (propositional or first-order)
logic, the denotation, or semantic value, of an expression is dependent solely
182 Appendix B

on the denotations of its sub-expressions. For example, the denotation of the

propositional logic formula p 1\ q is a function of the truth-values of p and q.
The operators of classical logic are thus said to be truthfunctional.
In contrast, intentional notions such as belief are not truth functional. It is
surely not the case that the truth value of the sentence:

Janine believes p (B.5)

l
is dependent solely on the truth-value of p. So substituting equivalents into
opaque contexts is not going to preserve meaning. This is what is meant by
referential opacity. The existence of referentially opaque contexts has been
known since the time of Frege. He suggested a distinction between sense and
reference. In ordinary formulae, the "reference" of a term/formula (i.e., its
denotation) is needed, whereas in opaque contexts, the "sense" of a formula
is needed (see also [204, p.3 ]).
Clearly, classical logics are not suitable in their standard form for reason­
ing about intentional notions: alternative formalisms are required. A vast en­
terprise has sprung up devoted to developing such formalisms.
The field of formal methods for reasoning about intentional notions is
widely reckoned to have begun with the publication, in 1962, of Jaakko Hin­
tikka's book K nowledge and Belief[96]. At that time, the subject was consid­
ered fairly esoteric, of interest to comparatively few researchers in logic and
the philosophy of mind. Since then, however, it has become an important re­
search area in its own right, with contributions from researchers in AI, formal
philosophy, linguistics, and economics. There is now an enormous literature on
the subject, and with a major biannual international conference devoted solely
to theoretical aspects of reasoning about knowledge, as well as the input from
numerous other, less specialized conferences, this literature is growing ever
larger [ 5 4].
Despite the diversity of interests and applications, the number of basic
techniques in use is quite small. Recall, from the discussion above, that there
are two problems to be addressed in developing a logical formalism for inten­
tional notions: a syntactic one, and a semantic one. It follows that any formal­
ism can be characterized in terms of two independent attributes: its language
offormulation, and semantic model[120, p.83 ].

1 Note, however, that the sentence (B.5) is itself a proposition, in that its denotation is the value
"true" or "false."
Formal Foundations 183

There are two fundamental approaches to the syntactic problem. The first
is to use a modal language, which contains non-truth-functional modal op­
erators, which are applied to formulae. An alternative approach involves the
use of a meta-language: a many-sorted first-order language containing terms
that denote formulae of some other object-language. Intentional notions can
be represented using a meta-language predicate, and given whatever axiomati­
zation is deemed appropriate. Both of these approaches have their advantages
and disadvantages, and will be discussed at length in what follows.
As with the syntactic problem, there are two basic approaches to the se­
mantic problem. The first, best-known, and probably most widely used ap­
proach is to adopt a possible worlds semantics, where an agent's beliefs,
knowledge, goals, etc. are characterized as a set of so-called possible worlds,
with an accessibility relation holding between them. Possible worlds seman­
tics have an associated correspondence theory, which makes them an attractive
mathematical tool to work with [3 1]. However, they also have many associated
difficulties, notably the well-known logical omniscience problem, which im­
plies that agents are perfect reasoners. A number of minor variations on the
possible worlds theme have been proposed, in an attempt to retain the corre­
spondence theory, but without logical omniscience.
The most common alternative to the possible worlds model for belief
is to use a sentential, or interpreted symbolic structures approach. In this
scheme, beliefs are viewed as symbolic formulae explicitly represented in a
data structure associated with an agent. An agent then believes <p if <p is present
in the agent's belief structure. Despite its simplicity, the sentential model works
well under certain circumstances [120].
The next part of this appendix contains detailed reviews of some of these
formalisms. First, the idea of possible worlds semantics is discussed, and then
a detailed analysis of normal modal logics is presented, along with some
variants on the possible worlds theme. Next, some meta-language approaches
are discussed, and one hybrid formalism is described. Finally, some alternative
formalisms are described.
Before the detailed presentations, a note on terminology. Strictly speaking,
an epistemic logic is a logic of knowledge, a doxastic logic is a logic of belief,
and a conative logic is a logic of desires or goals. However, it is common
practice to use "epistemic" as a blanket term for logics of knowledge and
belief. This practice is adopted in this book; a distinction is only made where
it is considered significant.
184 Appendix B

B.2 Possible Worlds Semantics

The possible worlds model for epistemic logics was originally proposed by
Hintikka [96], and is now most commonly formulated in a normal modal logic
using the techniques developed by Kripke [126].2
Hintikka's insight was to see that an agent's beliefs could be characterized
in terms of a set of possible worlds, in the following way. Consider an agent
playing the card game Gin Rummy.3 In this game, the more one knows about
the cards possessed by one's opponents, the better one is able to play. And
yet complete knowledge of an opponent's cards is generally impossible if one
excludes cheating. The ability to play Gin Rummy well thus depends, at least
in part, on the ability to deduce what cards are held by an opponent, given
the limited information available. Now suppose our agent possessed the ace of
spades. Assuming the agent's sensory equipment was functioning normally, it
would be rational of her to believe that she possessed this card. Now suppose
she were to try to deduce what cards were held by her opponents. This could
be done by first calculating all the various different ways that the cards in the
pack could possibly have been distributed among the various players. (This
is not being proposed as an actual card-playing strategy, but for illustration! )
For argument's sake, suppose that each possible configuration is described on
a separate piece of paper. Once the process was complete, our agent can then
begin to systematically eliminate from this large pile of paper all those con­
figurations which are not possible, given what she knows. For example, any
configuration in which she did not possess the ace of spades could be rejected
immediately as impossible. Call each piece of paper remaining after this pro­
cess a world. Each world represents one state of affairs considered possible,
given what she knows. Hintikka coined the term epistemic alternatives to de­
scribe the worlds possible given one's beliefs. Something true in all our agent's
epistemic alternatives could be said to be believed by the agent. For example,
it will be true in all our agent's epistemic alternatives that she has the ace of
spades.
On a first reading, this technique seems a peculiarly roundabout way
of characterizing belief, but it has two advantages. First, it remains neutral
on the subject of the cognitive structure of agents. It certainly doesn't posit

2 In Hintikka's original work, he used a technique based on "model sets," which is equivalent to
Kripke's formalism, though less elegant. See [101, Appendix Five, pp.351-352] for a comparison
and discussion of the two techniques.
3 This example was adapted from [88].
Formal Foundations 185

any internalized collection of possible worlds. It is just a convenient way of

characterizing belief. Second, the mathematical theory associated with the
formalization of possible worlds is extremely appealing (see below).
The next step is to show how possible worlds may be incorporated into the
semantic framework of a logic. This is the subject of the next section.

Normal Modal Logics

Epistemic logics are usually formulated as normal modal logics using the se­
mantics developed by Kripke [126]. Before moving on to explicitly epistemic
logics, this section describes normal modal logics in general.
Modal logics were originally developed by philosophers interested in the
distinction between necessary truths and mere contingent truths. Intuitively,
a necessary truth is something that is true because it could not have been
otherwise, whereas a contingent truth is something that could plausibly have
been otherwise. For example, it is a fact that as I write, the Labour Party of
Great Britain hold a majority in the House of Commons. But although this is
true, it is not a necessary truth; it could quite easily have turned out that the
Conservative Party won a majority at the last general election. This fact is thus
only a contingent truth.
Contrast this with the following statement: the square root of 2 is not a
rational number. There seems no earthly way that this could be anything but
true, given the standard reading of the sentence. This latter fact is an example
of a necessary truth. Necessary truth is usually defined as something true in all
possible worlds. It is actually quite difficult to think of any necessary truths
other than mathematical laws (those with strong religious views may well
disagree).
To illustrate the principles of modal epistemic logics, a normal proposi­
tional modal logic is defined.

Syntax and Semantics

This logic is essentially classical propositional logic, extended by the addition

of two operators: "0" (necessarily), and "<>" (possibly). First, its syntax.

DEFINITION B.l: Let Prop {p, q,...} be a countable set of atomic propo­
=

sitions. The syntax of normal propositional modal logic is defined by the fol­
lowing rules:

1. If P E Prop then p is a formula.

186 Appendix B

2. If cp,'lj; are formulae, then so are:

true -'cp cp V 'lj;

3 . If cp is a formula then so are:

The operators "-," (not) and "V" (or) have their standard meaning; true is a
logical constant, (sometimes called verum), that is always true. The remaining
connectives of propositional logic can be defined as abbreviations in the usual
way. The formula Dcp is read: "necessarily cp, " and the formula <) cp is read:
"possibly cp." Now to the semantics of the language.
Normal modal logics are concerned with truth at worlds; models for such
logics therefore contain a set of worlds, W, and a binary relation, R, on
W, saying which worlds are considered possible relative to other worlds.
Additionally, a valuation function 7f is required, saying what propositions are
true at each world.
A model for a normal propositional modal logic is a triple (W, R, 7f) , where
W is a non-empty set of worlds, R � W x W, and

7f : W --t p( Prop)

is a valuation function, which says for each world w E W which atomic

propositions are true in w. An alternative, equivalent technique would have
been to define 7f as follows:

7f: W x Prop --t {true, false}

though the rules defining the semantics of the language would then have to be
changed slightly.
The semantics of the language are given via the satisfaction relation, "F , "
which holds between pairs of the form (M,w) (where M is a model, and w is
a reference world), and formulae of the language. The semantic rules defining
this relation are given in Figure B.I. The definition of satisfaction for atomic
propositions thus captures the idea of truth in the "current" world, which
appears on the left of " F ". The semantic rules for "true," "-'," and "V" are
standard. The rule for " D" captures the idea of truth in all accessible worlds,
and the rule for "<)" captures the idea of truth in at least one possible world.
Formal Foundations 187

(M,w) F= true
(M,w) F= p wherep E Prop, iffp E 7r(w)
(M,w) F= -''P iff (M,w) fto 'P
(M,w) F= 'PV'lj; iff (M,w) F= 'P or (M,w) F= 'Ij;
(M,w) F= D'P iff\iw' E W· if (w,w') E Rthen (M,w') F= 'P
(M,w) F= O'P iff:Jw' E W· (w,w') E Rand (M,w') F= 'P

Figure B.1
The semantics of normal modal logic.

Note that the two modal operators are duals of each other, in the sense that
the universal and existential quantifiers of first-order logic are duals:

It would thus have been possible to take either one as primitive, and introduce
the other as a derived operator.

Correspondence Theory

To understand the extraordinary properties of this simple logic, it is first

necessary to introduce validity and satisfiability. A formula is satisfiable if it
is satisfied for some model/world pair, and unsatisfiable otherwise. A formula
is true in a model if it is satisfied for every world in the model, and valid in a
class of models if it true in every model in the class. Finally, a formula is valid
simpliciter if it is true in the class of all models. If cp is valid, we write F cpo
The two basic properties of this logic are as follows. First, the following
axiom schema is valid.

This axiom is called K, in honor of Kripke. The second property is as follows.

if F cp then F Dcp

Proofs of these properties are trivial, and are left as an exercise for the reader.
Now, since K is valid, it will be a theorem of any complete axiomatization
of normal modal logic. Similarly, the second property will appear as a rule
of inference in any axiomatization of normal modal logic; it is generally
called the necessitation rule. These two properties turn out to be the most
problematic features of normal modal logics when they are used as logics of
knowledgelbelief (this point will be examined later).
188 Appendix B

Table B.1
Some correspondence theory.

Axiom Condition First-Order

Name Schema on R Characterization

T D<p => <p reflexive Vw E W· (w,w) E R

D D<p => <><p serial Vw E W· :Jw' E W· (w,w') E R

4 D<p => D D<p transitive VW,W',W" E W· (w,w/ ) E R /\

(W',W") E R => (w,w" ) E R

5 <><p => D <><p Euclidean VW,W',W" E W· (w,w/ ) E R /\

(w,w" ) E R => (w',w" ) E R

The most intriguing properties of normal modal logics follow from the
properties of the accessibility relation, R, in models. To illustrate these proper­
ties, consider the following axiom schema.

Dcp=>cp

It turns out that this axiom is characteristic of the class of models with a
reflexive accessibility relation. (By characteristic, we mean that it is true in
all and only those models in the class.) There are a host of axioms that
correspond to certain properties of R: the study of the way that properties of
R correspond to axioms is called correspondence theory. In Table B.l, we list
some axioms along with their characteristic property on R, and a first-order
formula describing the property. Note that the table only lists those axioms of
specific interest to this book; see [3 1] for others. The names of axioms follow
historical tradition.
The results of correspondence theory make it straightforward to derive
completeness results for a range of simple normal modal logics. These results
provide a useful point of comparison for normal modal logics, and account in
a large part for the popularity of this style of semantics.
A system of logic can be thought of as a set of formulae valid in some
class of models; a member of the set is called a theorem of the logic (if cp is
a theorem, this is usually denoted by f- cp). The notation K�l . . . �n is often
used to denote the smallest normal modal logic containing axioms �l' . . . , �n
(any normal modal logic will contain K; cf. [81, p.25]).
For the axioms T, D, 4, and 5, it would seem that there ought to be sixteen
distinct systems of logic (since 24 16). However, some of these systems tum
=

out to be equivalent (in that they contain the same theorems), and as a result
Formal Foundations 189

K4 K5 KD ---_I KT=KDT

K45 KD5 KD4 ---_I KT4=KDT4

�/ KD45
KT45
--------. KT5= KDT5
KDT45

Figure B.2
The normal modal logic systems based on axioms T, D, 4, and 5.

there are only eleven distinct systems. The relationships among these systems
are described in Figure B.2 (after [120, p.99], and [3 1, p.13 2]). In this diagram,
an arc from A to B means that B is a strict superset of A: every theorem of A is
a theorem of B, but not vice versa ; A B means that A and B contain precisely
=

the same theorems.

Because some modal systems are so widely used, they have been given
names:

• KT is known as T;
• KT4 is known as S4;
• KD45 is known as weak-S5; and
• KT5 is known as S5.

Normal Modal Logics as Epistemic Logics

To use the logic developed above as an epistemic logic, the formula D<p is read
as: "it is known that <p." The worlds in the model are interpreted as epistemic
190 Appendix B

alternatives, the accessibility relation defines what the alternatives are from
any given world. The logic deals with the knowledge of a single agent. To deal
with multiagent knowledge, one adds to a model structure an indexed set of
accessibility relations, one for each agent. A model is then a structure:

where R; is the knowledge accessibility relation of agent i. The simple language

defined above is extended by replacing the single modal operator " D" by an
indexed set of unary modal operators {K;}, where i E {I, . . . ,n }. The formula
K;cp is read: "i knows that cp. " The semantic rule for " D" is replaced by the
following rule:

(M,w) F K;cp iff\fw' E W· if ( w, w' ) E R; then (M,w' ) F cp

Each operator K;thus has exactly the same properties as D." Corresponding
"

to each of the modal systems �, above, a corresponding system �n is defined,

for the multiagent logic. Thus Kn is the smallest multiagent epistemic logic
and S5n is the largest.
The next step is to consider how well normal modal logic serves as a logic
of knowledgelbelief. Consider first the necessitation rule and axiom K, since
any normal modal system is committed to these.
The necessitation rule tells us that an agent knows all valid formulae.
Among other things, this means an agent knows all propositional tautologies.
Since there are an infinite number of these, an agent will have an infinite
number of items of knowledge: immediately, one is faced with a counter­
intuitive property of the knowledge operator.
Now consider the axiom K, which says that an agent's knowledge is
closed under implication. Suppose cp is a logical consequence of the set � =

{CPI, . . . ,CP n }, then in every world where all of � are true, cp must also be true,
and hence

CPI 1\ ... 1\ CP n=>cP

must be valid. By necessitation, this formula will also be believed. Since an

agent's beliefs are closed under implication, whenever it believes each of �,
it must also believe cpo Hence an agent's knowledge is closed under logical
consequence. This also seems counter-intuitive. For example, suppose, like
every good logician, our agent knows Peano's axioms. Now Fermat's last
theorem follows from Peano's axioms - but it took several hundred years,
Formal Foundations 19 1

and an unthinkable amount of hard work before Andrew Wiles was able to
prove it. But if our agent's beliefs are closed under logical consequence, then
our agent must know it. So consequential closure, implied by necessitation and
the K axiom, seems an overstrong property for resource-bounded reasoners.

The Logical Omniscience Problem

These two problems - that of knowing all valid formulae, and that of know1-
edgelbelief being closed under logical consequence - together constitute the
famous logical omniscience problem. This problem has some damaging corol­
laries.
The first concerns consistency. Human believers are rarely consistent in the
logical sense of the word; they will often have beliefs rp and 'ljJ, where rp f- -,'ljJ,
without being aware of the implicit inconsistency. However, the ideal reason­
ers implied by possible worlds semantics cannot have such inconsistent beliefs
without believing every formula of the logical language (because the conse­
quential closure of an inconsistent set of formulae is the set of all formulae).
Konolige has argued that logical consistency is much too strong a property
for resource-bounded reasoners: he argues that a lesser property - that of be­
ing non-contradictory - is the most one can reasonably demand [120]. Non­
contradiction means that an agent would not simultaneously believe rp and -'rp,
although the agent might have logically inconsistent beliefs.
The second corollary is more subtle. Consider the following propositions
(this example is from [120, p.88]):

1. Hamlet's favorite color is black.

2. Hamlet's favorite color is black and every planar map can be four colored.

The second conjunct of (2) is valid, and will thus be believed. This means
that (1) and (2) are logically equivalent; (2) is true just when (1) is. Since agents
are ideal reasoners, they will believe that the two propositions are logically
equivalent. This is yet another counter-intuitive property implied by possible
worlds semantics, as: "equivalent propositions are not equivalent as beliefs"
[120, p.88]. Yet this is just what possible worlds semantics implies. It has been
suggested that propositions are thus too coarse-grained to serve as the objects
of belief in this way.
The logical omniscience problem is a serious one. In the words of
Levesque:

Any one of these [problems] might cause one to reject a possible-world formalization
19 2 Appendix B

as unintuitive at best and completely unrealistic at worst. [13 2]

Axioms for Knowledge and Belief

We now consider the appropriateness of the axioms Dn, Tn, 4n, and 5n for
logics of knowledgelbelief.
The axiom Dn says that an agent's beliefs are non-contradictory; it can be
rewritten in the following form:

which is read: "if i knows <p, then i doesn't know '<p." This axiom seems a
reasonable property of both knowledge and belief.
The axiom Tn is often called the knowledge axiom, since it says that what is
known is true. It is usually accepted as the axiom that distinguishes knowledge
from belief: it seems reasonable that one could believe something that is false,
but one would hesitate to say that one could know something false. Knowledge
is thus often defined as true belief: i knows <p if i believes <p and <p is true. So
defined, knowledge satisfies Tn.
Axiom 4n is called the positive introspection axiom. Introspection is the
process of examining one's own beliefs, and is discussed in detail in [120,
Chapter S]. The positive introspection axiom says that an agent knows what it
knows. Similarly, axiom 5n is the negative introspection axiom, which says that
an agent is aware of what it doesn't know. Positive and negative introspection
together imply an agent has perfect knowledge about what it does and doesn't
know (cf. [120, Equation (S.l1), p.79]). Whether or not the two types of
introspection are appropriate properties for knowledgelbelief is the subject
of some debate. However, it is generally accepted that positive introspection
is a less demanding property than negative introspection, and is thus a more
reasonable property for resource-bounded reasoners.
Given the comments above, the modal system SSn is often chosen as a
logic of knowledge, and weak-SSn is often chosen as a logic of belief

Computational Aspects

Before leaving this basic logic, it is worth commenting on its computa­

tional/proof theoretic properties. Halpern and Moses have established the fol­
lowing ([90]):

1. The satisfiability problem for each of the key systems Kn, Tn, S4n, weak­
SSn, and SSn is decidable.
Formal Foundations 193

2. The satisfiability and validity problems for Kn, Tn, S4n (where n � 1), and
S5n, weak-S5n (where n � 2) are P SPACE complete.

The first result is encouraging, as it holds out at least some hope of automation.
Unfortunately, the second result is extremely discouraging: in simple terms,
it means that in the worst case, automation of these logics is not a practical
proposition.

Discussion

To sum up, the basic possible worlds approach described above has the follow­
ing disadvantages as a multiagent epistemic logic:

• agents believe all valid formulae;

• agents' beliefs are closed under logical consequence;
• equivalent propositions are identical beliefs;
• if agents are inconsistent, then they believe everything;
• in the worst case, automation is not feasible.

To which many people would add the following:

The ontology of possible worlds and accessibility relations . . . is frankly mysterious

to most practically minded people, and in particular has nothing to say about agent
architecture. [204]

Despite these serious disadvantages, possible worlds are still the semantics
of choice for many researchers, and a number of variations on the basic possi­
ble worlds theme have been proposed to get around some of the difficulties.

Quantified Epistemic Logics

A natural extension to the basic logic described above would be to allow

quantification. In this section, we proceed (as with the propositional modal
logic described above) to define a quantified modal logic and then examine
its suitability as a logic for reasoning about knowledge/belief. The discussion
below is adapted and expanded from [101, Part Two], [192], and [120].
The syntax of quantified modal logic is that of classical first-order logic
enriched by the addition of the unary modal operators <> and D, with
" " " "

the same meanings as above. These operators can be applied to arbitrary

formulae of the modal or first-order language. So, for example, the following
194 Appendix B

are syntactically acceptable formulae of quantified modal logic:

Vx· DP(x) => Q(x)

P(a) 1\ D3x· P(x)
It is assumed that constants of the basic first-order language belong to a set
Canst, variables to a set Var, and predicate symbols to a set Pred. A model for
quantified modal logic is a structure:

(W, R, D, I, 7r)
where W and R are a set of worlds and a binary relation on W (as before), and:

• D is a non-empty set, called the domain ;

• I maps Canst to D;
• 7r maps Pred x W to p(IY'), where n is the arity of the predicate symbol,

and IY' is the set of n-tuples over D; the function 7r gives the extension of each
predicate symbol in each world.

As usual with a predicate logic, formulae can only be interpreted with

respect to a variable assignment: a variable assignment V maps Var to D. A
term is either a variable or a constant. It is useful to define a function [ . . .]J,v,
which takes an arbitrary term and returns its denotation relative to I, V :

ifT E Canst
otherwise.

Reference to I and V is suppressed where these are understood.

The only semantic rule of interest is that for predicates; the remainder are
essentially unchanged, except that a formula must now be interpreted with
respect to a variable assignment.

So, despite the added complexity of predicates and terms, the model theory for
quantified modal logics remains straightforward.
Now consider such a system as an epistemic logic. The semantics can
easily be extended to the multiagent case, as in propositional modal logic, by
the addition of multiple accessibility relations, etc. One unfortunate property
of quantified modal logics, as defined above, is that they make the Barcan
formulae valid. The Barcan formula is:

Vx· Dcp(x) => DVx· cp(x)

Formal Foundations 195

The converse Barcan formula is the reverse implication:

D\lx· cp(x) ::::} \Ix· Dcp(x)

(see e.g., [101, pp.142-145] for a discussion).

If " D" is interpreted as "it is known that . . . ," then the Barcan formula
means that if it is known of all individuals independently that they have
property cp, then it is known that all individuals have this property. The Barcan
formula thus implies that knowers are somehow aware of the existence of all
individuals, which seems to be an overstrong demand to make of resource­
bounded reasoners.
The converse Barcan formula says that if an agent knows that all individ­
uals have property cp, then it will be known that each individual has property
cpo In other words, it implies that agents will perform universal instantiation
wherever possible. This is a weaker property than that implied by the Barcan
formula, but still seems an overstrong demand to make of resource-bounded
reasoners.
If one refers back to model structures for quantified modal logics for a
moment, it will be found that the domain of individuals D is fixed, for all
worlds: this is called the constant domain assumption. If this assumption
is dropped, so that the domain is indexed by worlds, and quantification is
restricted to just those individuals in the domain of the reference world, then
the Barcan formulae are no longer valid. As to whether this is a reasonable
property for an epistemic logic, consider a semi-mythological figure such as
Robin Hood, or King Arthur. An agent might reasonably be unsure as to
whether these individuals ever actually existed, and so they might appear in
the domain of some epistemic alternatives, but not all.
It is also worth noting that in the model structure above, constants were
given a rigid interpretation: a constant has the same denotation in all worlds.
What would it mean for this assumption to be dropped, and what would be the
effect? It would be possible to drop the assumption by simply making I a map
from Canst x W to D. Consider the following formula, which is intended to
express the fact that an agent believes of the individual that is Prime Minister
of Britain in the actual world - the world as perceived by the author at the
time of writing - that he is a cricket fan.

DCricketFan( PM)

Here, PM is a constant which in the actual world denotes the individual Tony
196 Appendix B

Blair (Prime Minister of the U.K. at the time of writing). But suppose our
agent had been on a desert island since November 1990, and was somehow
unaware that Margaret Thatcher was no longer Prime Minister. For this agent,
the constant PM would denote someone entirely different to that which it
denotes for us, and it is unlikely that a rational agent would believe of the
individual Margaret Thatcher that she was a cricket fan. Note that the agent
could still believe that the individual Tony Blair was a cricket fan, but in its
ignorance, it would use a different constant for this individual. The point is
that the constant PM has a different denotation for us and the agent, and, more
generally, a different denotation in different worlds. Such constants are said to
befluent expressions; non-fluent expressions are called rigid designators.
To deal with fluent expressions, Konolige has proposed the technical de­
vice of the bullet operator, ("e" ) . Suppose a is a constant, then ea is an expres­
sion returning a constant that always denotes what a denotes for us. The use
of the bullet operator is somewhat complex, and is described in detail in [120,
pp.3 8--42 and pp.l00-104].
One difficult issue in the philosophy of modal logics is quantifYing-in to
modal contexts. Consider the following English sentence:

It is known that there is a unicorn. (B.6)

Now consider the following quantified modal logic formulae, each represent­
ing one attempt to formalize (B.6):

3x . DUnicorn(x) (B.7)
D:3x· Unicorn(x) (B.8)

In (B.7), a de re reading of (B.6) is presented. The formula implies the exis­

tence of some particular individual, which the believer has a name for, which
the agent believes has the property of being a unicorn. Note that we do not
know who this individual is, or have a name for it; we simply represent the fact
that the agent does.
In (B.8), a de dicto reading of (B.6) is presented. This reading expresses
a weaker property than (B. 7): namely, that the agent believes some individual
is a unicorn. The agent does not necessarily have a name for the individual.
Formulae (B.7) and (B.8) thus express quite different properties.
Formal Foundations 19 7

Meta-Languages and Syntactic Modalities

There is a school of thought - almost a cult - in mathematics, computer

science, and AI, which holds that classical first-order logic is in some sense
"canonical" : that anything which can be done in a non-classical (e.g., modal)
logic can be reduced to a problem in first-order logic. In the area of reasoning
about intentional notions, this thinking manifests itself in the use of first-order
meta-languages to represent and reason about beliefs, goals, etc.
The basic idea of a meta-language is quite simple. Recall the naive attempt
to formalize the statement "Janine believes Cronos is the father of Zeus"
presented earlier:

Bel(Janine, Father (Zeus, Cronos) ) (B.9)

The syntactic problem with this formalization is that Father(Zeus, Cronos) is

a formula of first-order logic, not a term, so the whole formula is ill-formed.
Now suppose the domain of the first-order language (hereafter called the meta­
language) contains formulae of another language (hereafter called the object­
language). Suppose also that the meta-language contains terms which denote
object-language formulae. Then it would be possible to write a meta-language
formula capturing the sense of (B.9).
To make meta-language formulae readable, it is useful to employ a quoting
convention. Suppose that 'P is an object-language formula, then it is usual to let
I'P1 be an abbreviation for a meta-language term standing for 'P. A formulation
of the above sentence might then be:

Bel(Janine,I Father (Zeus, Cronos) 1) (B.IO)

Since IFather(Zeus, Cronos) l is a meta-language term, (B.IO) is a syntacti­

cally acceptable formula of the meta-language. Using this approach, belief,
knowledge, etc. can be axiomatized using the first-order meta-language and
given whatever properties are deemed appropriate. One obvious difficulty
seems to be that it is not possible to quantify over object-language terms.
This difficulty can be overcome by using an un-quoting convention: see [73 ,
pp.241-242] for a discussion of this point.
The quote marks, 1 ...1, are sometimes called Frege quotes, or sense
quotes. The process of making formulae into objects in a meta-language do­
main is sometimes called arithmetization, after G6del, who used a meta­
language technique in his proof of the incompleteness of arithmetic; the names
given to object-language formulae are sometimes called Godel numbers.
198 Appendix B

It is often claimed that meta-language approaches enjoy the following

advantages over modal languages:

1. Expressive power.
The following meta-language formulae have no quantified modal logic coun­
terparts:
:3x . Bel(i, x) "i believes something"
\:Ix· Bel(i, x) ::::} BeIU, x) "i believes everything) believes"
2. Computational tractability.
Meta-languages are just many-sorted first-order languages, for which auto­
mated theorem provers exist. It should therefore be possible - in principle
at least - to use existing theorem provers for meta-languages. However, this
claim has yet to be satisfactorily demonstrated.

The key difficulty with meta-language approaches is that many naive at­
tempts to treat meta-language predicates as syntactic modalities run into in­
consistency. These difficulties arise, for the large part, from the issue of self
reference. 4
Suppose one wished to use a meta-language approach to represent and
reason about knowledge and belief. Following "standard" practice, knowledge
may be defined as true belief. One might begin by setting object-language =

meta-language, so that the object- and meta-language are the same; in such
circumstances, the meta-language is said to be self-referential, since it may
contain terms that refer to its own formulae. A truth predicate can then be
defined.

(B.lI)

Knowledge may then be defined as a composite concept.

(B.12)

The Bel predicate can then be given whatever axiomatization is deemed appro­
priate. This approach seems simple, intuitive, and satisfactory. Unfortunately,
in general, any first-order meta-language theory containing axiom (B.II) turns
out to be inconsistent. This problem was first recognized by Tarski; the diffi­
culty is with the possibility of formulae asserting their own falsity (as in the
famous "liar" paradox).

4 A good discussion of these issues may be found in [173]; the issue of self reference is examined
in detail in [171] and [172].
Formal Foundations 199

Several proposals have been put forward to remedy the problem. One
possibility is to define True and False predicates, such that for some formulae
1 1
<p, neither True( r <p ) nor False( r <p ) hold; in effect, one axiomatizes a system
in which the law of the excluded middle does not hold. This solution is
unintuitive, however, and seems rather ad hoc.
Another solution, due to Perlis [171], is to replace (B.ll) with an axiom of
the form:

(B.13 )
*
where <p is the formula obtained from <p by replacing every occurrence of
-, True(r 'lj; 1) in <p by True(r -,'lj; 1) . This simple expedient prevents inconsistency
(see [171, pp.3 12-315] for proof). This proposal has been criticized by Kono­
lige, on a number of grounds, however: see [120, p.116] for details.
A problem not alleviated by Perlis' schema was first recognized by Mon­
tague [155]. Consider the following axiom:

and the following inference rule:

1
From f-- <p infer f-- Bel(r <p )

(cf. the modal system KT). Montague showed that any moderately sophis­
ticated first-order theory (and in particular, basic arithmetic) containing the
above axiom and inference rule is inconsistent. This result appears to be dev­
astating, and for a long time dampened down research into first-order meta­
languages. A similar result by Thomason ([229]) is also discouraging.
However, des Rivieres and Levesque have recently shown that, while
Montague's results are technically correct, a careful reworking, with slightly
different assumptions, leads back to consistency [45].5 Also, Perlis has shown
how a similar technique to that proposed by him for recovering from Tarski's
result can be used to recover from Montague's results [172].
If nothing else, the above discussion illustrates that the whole issue of self­
referential languages is a difficult one. It is hardly surprising, therefore, that
some researchers have rejected self reference entirely. One critic is Konolige,
who suggests that an observer of a system of agents can hardly have any use
for a self-referential language. He argues that the logical puzzles caused by

5 In fact, they showed that it was a naive translation from a modal to meta-language that was
causing difficulties.
200 Appendix B

self-reference are just that: puzzles, which an agent will never usually need to
consider (unless she is a philosophy or logic student) [120, p.llS].
One "standard" alternative to self-referential languages is to use an hierar­
chicallanguage structure. Consider a "tower" of languages:

Lo - L1 - L2 - ... - Lk - ...

Let language Lo be a non-self-referential "base" language; standard first-order

logic, for example. Then let each of the languages Lk, where k > 0, be a
meta-language containing terms referring only to formulae of languages lower
down in the hierarchy. Thus no language in the hierarchy is self-referential. It
is possible to write formulae such as (B.l0) in any of the languages Lk, where
k > 0, but the paradoxes of self-reference will not arise.
The hierarchical approach has problems of its own, however. Consider the
following scenario, adapted from [23 1]. There are two agents, A and B, each
with equal expressive powers of belief: anything A can believe, B can believe,
and vice versa. Consider the following statement.

A believes that everything B believes is true. (B.14)

How might this statement be represented, using the hierarchical approach? It

obviously cannot be represented in Lo. One obvious solution is to write:

Bel(A,r \/x· Bel(B,x) => True(x) 1).

Suppose that the inner Bel predicate belongs to language Lk, where k > o.
The variable x in this formula ranges over all formulae of languages L}> where
j < k. It does not, therefore, range over the language of B's beliefs, but over
some language lower in the hierarchy. Moving to a language farther up the
hierarchy does not solve the problem:

No matter how far we climb up the object/meta-language hierarchy, we will not be able
to capture the intuitive content of [(B. 14)]. [23 1, pp. 7-8]

Other objections to the hierarchical approach have been raised. For exam­
ple, Morgenstern points out that the truth predicate paradox is only resolved
by positing an infinite number of truth predicates, one for each language in the
hierarchy. She argues that this is not a satisfactory solution, for two reasons:

In the first place, it is implausible that people are consciously aware of using different
truth predicates when they speak. Secondly, we often do not know which truth predicate
to use when we utter a particular statement. Finally, it is impossible . . . to construct a
pair of sentences, each of which refers to the other, although such sentences may make
perfect sense. [160, p. 104]
Formal Foundations 20 1

So both self-referential and hierarchical languages have their difficulties.

One problem common to both is that meta-languages are, in general, notation­
ally cumbersome, unintuitive, and complex: see [120, p.lll] for a discussion
of this point. However, the difficulties described above have not prevented the
development of some ingenious and useful meta-language formulations of in­
tentional notions. Examples include [119, 41].

B.3 Temporal Logics

There seem to be almost as many different temporal logics as there are people
using temporal logics to reason about programs. In this section, we briefly
describe some important developments in temporal logics.
The earliest temporal logic studied in detail was the modal logic Kt, cor­
responding to a modal system with two basic modal operators (often written
"[P]" or "H" - "heretofore," or "always past," and "[F ] " or "G" - "hence­
forth," or "always") [182]. The semantics of these operators are given via two
ordering relations Rp and RF on the set of worlds, such that the two relations
are the inverse of each other. The usual "was" and "sometime" operators (writ­
ten "( P)" or " P " and "(F ) " or "F") are defined as the duals of these operators.
This logic is discussed at length in [81, Chapter 6].
Although such a system can be used to crudely reason about programs, it
cannot describe the "fine structure" of the state sequences generated by exe­
cuting programs (see below). For this reason, a "next time" operator was used
by Pnueli in his original proposal [176]. Also, the standard heretofore/always
combination of operators was discovered to be inadequate for expressing many
properties, and so the basic temporal logic was augmented by the since and un­
til operators (written " S" and " U "). In his 1968 doctoral thesis, Kamp demon­
strated that the logic with since and until operators was expressively complete
over continuous linear orders [115].6
The "standard" linear discrete temporal logic is based on until/since and
next/last [144, 145]. In the propositional case, this logic is decidable (though
the problem is P SPACE complete - see [219]), and is generally regarded as
expressive enough to capture many interesting properties of reactive systems.
Gough has developed tableaux-based decision procedures for this logic [82];
Fisher has developed resolution methods [62].

6 The since/until operators are assumed to be strict: the since operator refers strictly to the past,
the until operator refers strictly to the future.
20 2 Appendix B

In the first-order case, temporal logics containing these operators are not
decidable, but this should come as no surprise! Perhaps more worrying is that
first-order temporal logic is not finitely axiomatizable [1].
A temporal logic based solely on these operators does have its limitations,
however. Sistla et al. showed that it could not be used to describe an unbounded
FIFO buffer [221].
Many variations on this basic temporal logic theme exist. For example, the
use offixpoint operators has been discussed. Banieqbal and Barringer describe
a logic that keeps the next operator as basic, but then defines two fixpoint
operators from which the other standard operators of linear discrete temporal
logic can be derived [ 7]. Wolper has described a logic ETL, which extends the
standard temporal logic with a set of grammar operators [23 5]. Yet another
variation is to add a chop operator, C, which "composes" two finite sequences
[9].
All of the logics mentioned above have assumed the model of time is
discrete (i.e., that for any time point, there exists a time such that no other
time point occurs between them) and linear (i.e., that each point in time has at
most one successor). But these assumptions are not essential: a temporal logic
of reals (where the worlds are isomorphic with the real numbers) has been
developed; however, such logics are complex and uncommon.
Much more common are branching time logics. Briefly, a branching time
structure is one where each time point may have a number of "successor"
times, each one intuitively corresponding to one way things could turn out.
The repertoire of operators in linear temporal logic is not sufficient to express
all the properties of such structures (despite Lamport's claims in his famous
1980 paper [129]).
The earliest branching time logic to be studied at length was UB (unified
system of branching time) [12]. This logic extended linear time temporal logic
by the addition of two operators, "A" ("on all paths . . . ") and "E" ("on some
path . . . "), called path quantifiers, which could be prefixed to any formula
containing at most one occurrence of the usual linear time temporal operators.
This introduces an obvious restriction on the expressibility of UB.
UB does not contain an "until" operator; this omission was rectified in a
logic CTL (computation tree logic). However, CTL is still not expressive enough
to capture many interesting properties of branching time structures. The most
expressive branching time logic so far studied in any detail is called CTL * [50].
This logic allows path quantifiers to be prefixed to arbitrary formulae of the
language: intermixing of path quantifiers and standard temporal logic operators
Formal Foundations 203

is freely allowed, resulting in a highly expressive logic. In the propositional

case, all of these branching time logics are known to be decidable. A good
general review of branching time logics is provided in [51]. CTL * forms the
basis of CORA, the logic used in the main text of this book.
References

[1] M. Abadi. Temporal Logic Theorem Proving. PhD thesis, Computer Science
Department, Stanford University, Stanford, CA 94305, 1987.

[2] J. F. Allen, 1. Hendler, and A. Tate, editors. Readings in Planning. Morgan

Kaufmann Publishers: San Mateo, CA, 1990.

[3] D. E. Appelt. Planning natural language utterances. In Proceedings of the

Second National Conforence on Artificial Intelligence (AAAI-82), pages 59-62,
Pittsburgh, PA, 1982.

[4] D. E. Appelt. Planning English Sentences. Cambridge University Press: Cam­

bridge, England, 1985.

[5] D. E. Appelt and K. Konolige. A nonmonotonic logic for reasoning about speech
acts and belief revision. In M. Reinfrank, 1. de Kleer, M. L. Ginsberg, and
E. Sandewall, editors,
Non-Monotonic Reasoning - Proceedings of the Second
International Workshop (LNAI Volume 346), pages 164-175. Springer-Verlag:
Berlin, Germany, 1988.

[6] J. L. Austin. How to Do Things With Words. Oxford University Press: Oxford,
England, 1962.

[7] B. Banieqbal and H. Barringer. A study of an extended temporal language and a

temporal fixed point calculus. Technical Report UMCS-86-10-2, Department of
Computer Science, Manchester University, Oxford Rd. , Manchester M13 9PL,
UK, 1986.

[8] H. Barringer, M. Fisher, D. Gabbay, G. Gough, and R. Owens. METATEM: A

framework for programming in temporal logic. In REX Workshop on Stepwise
Refinement of Distributed Systems: Models, Formalisms, Correctness (LNCS
Volume 430), pages 94-129. Springer-Verlag: Berlin, Germany, June 1989.
[9] H. Barringer, R. Kuiper, and A. Pnueli. Now you may compose temporal logic
specifications. In Proceedings of the Sixteenth ACM Symposium on the Theory
of Computing, pages 51-63, 1984.

[10] H. Barringer, R. Kuiper, and A. Pnueli. A really abstract concurrent model and
its temporal logic. In Proceedings of the Thirteenth ACM Symposium on the
Principles of Programming Languages, pages 173-183, 1986.
[11] M. Ben-Ari. Principles of Concurrent and Distributed Programming. Prentice
Hall, 1990.

[12] M. Ben-Ari, Z. Manna, and A. Pnueli. The temporal logic of branching time. In
Proceedings of the Eighth ACM Symposium on the Principles of Programming
Languages (POPL), pages 164-176, 1981.
[13] M. Benerecetti, F. Giunchiglia, and L. Serafini. A model checking algorithm
for multiagent systems. In J. P. MUller, M. P. Singh, and A. S. Rao, editors,
Intelligent Agents V (LNAI Volume 1555). Springer-Verlag: Berlin, Germany,
1999.

[14] K. Binmore. Fun and Games: A Text on Game Theory. D. C. Heath and
Company: Lexington, MA, 1992.
206 References

[15] L. Birnbaum. Rigor mortis. In D. Kirsh, editor, Foundations of Artificial

Intelligence, pages 57-78. The MIT Press: Cambridge, MA 1992. ,

[16] J. Blythe. An overview of planning under uncertainty. In M. Wooldridge and

M. Veloso, editors, Artificial Intelligence Today (LNAI 1600), pages 85-110.
Springer-Verlag: Berlin, Germany, 1999.

[17] R. P. Bonasso, D. Kortenkamp, D. P. Miller, and M. Slack. Experiences with an

architecture for intelligent, reactive agents. In M. Wooldridge, J. P. Miiller, and
M. Tambe, editors, Intelligent Agents II (LNAI Volume 1037), pages 187-202.
Springer-Verlag: Berlin, Germany, 1996.

[18] A. H. Bond and L. Gasser, editors. Readings in Distributed Artificial Intelli­

gence. Morgan Kaufmann Publishers: San Mateo, CA, 1988.

[19] G. Booch. Object-Oriented Analysis and Design (second edition). Addison­

Wesley: Reading, MA 1994. ,

[20] M. E. Bratman. Intention, Plans, and Practical Reason. Harvard University

Press: Cambridge, MA 1987. ,

[21] M. E. Bratman. W hat is intention? In P. R. Cohen, J. L. Morgan, and M. E.

Pollack, editors, Intentions in Communication, pages 15-32. The MIT Press:
Cambridge, MA 1990. ,

[22] M. E. Bratman, D. J. Israel, and M. E. Pollack. Plans and resource-bounded

practical reasoning. Computational Intelligence, 4:349-355, 1988.

[23] P. Bretier and D. Sadek. A rational agent as the kernel of a cooperative spoken
dialogue system: Implementing a logical theory of interaction. In J. P. Miiller,
M. Wooldridge, and N. R. Jennings, editors, Intelligent Agents III (LNAI Volume
1193), pages 189-204. Springer-Verlag: Berlin, Germany, 1997.

[24] R. A. Brooks. Cambrian Intelligence. The MIT Press: Cambridge, MA 1999. ,

[25] M. A. Brown. On the logic of ability. Journal of Philosophical Logic, 17:1-26,

1988.

[26] P. Busetta, N. Howden, R. Ronnquist, and A. Hodgson. Structuring BD! agents

in functional clusters. In N.R. Jennings and Y. Lesperance, editors, Intelligent
Agents VI - Proceedings of the Sixth International Workshop on Agent Theo­
ries, Architectures, and Languages (ATAL-99), Lecture Notes in Artificial Intel­
ligence, pages 277-289. Springer-Verlag, Berlin, 2000.

[27] C. Castelfranchi. Social power. In Y. Demazeau and J. -P. Miiller, editors,De­

centralized AI - Proceedings of the First European Workshop on Modelling Au­
tonomous Agents in a Multi-Agent World (MAAMAW-89), pages 49-62. Elsevier
Science Publishers B. Y.: Amsterdam, The Netherlands, 1990.

[28] C. Castelfranchi. Guarantees for autonomy in cognitive agent architecture. In

M. Wooldridge and N. R. Jennings, editors, Intelligent Agents: Theories, Ar­
chitectures, and Languages (LNAI Volume 890), pages 56-70. Springer-Verlag:
Berlin, Germany, January 1995.
References 207

[29] L. Catach. Normal multimodal logics. In Proceedings of the Seventh National

Coriference on Artificial Intelligence (AAAI-88), pages 491-495, St. Paul, MN,
1988.

[30] L. Catach. TABLEAUX: A general theorem prover for modal logics. Journal of
Automated Reasoning, 7:489-510, 1991.

[31] B. Chellas. Modal Logic: An Introduction. Cambridge University Press: Cam­

bridge, England, 1980.

[32] E. M. Clarke and E. A. Emerson. Design and synthesis of synchronization

skeletons using branching time temporal logic. In D. Kozen, editor, Logics of
Programs - Proceedings 1981 (LNCS Volume 131), pages 52-7l. Springer­
Verlag: Berlin, Germany, 1981.

[33] E. M. Clarke, o. Grumberg, and D. A. Peled. Model Checking. The MIT Press:
Cambridge, MA 2000.,

[34] W. F. Clocksin and C. S. Mellish. Programming in Prolog. Springer-Verlag:

Berlin, Germany, 1981.

[35] P. R. Cohen and H. 1. Levesque. Intention is choice with commitment. Artificial

Intelligence, 42:2l3-261, 1990.

[36] P. R. Cohen and H. 1. Levesque. Rational interaction as the basis for commu­
nication. In P. R. Cohen, 1. Morgan, and M. E. Pollack, editors, Intentions in
Communication, pages 221-256. The MIT Press: Cambridge, MA 1990. ,

[37] P. R. Cohen and H. J. Levesque. Teamwork. Nous, 25(4):487-512, 1991.

[38] P. R. Cohen and H. 1. Levesque. Communicative actions for artificial agents.

In Proceedings of the First International Conference on Multi-Agent Systems
(ICMAS-95), pages 65-72, San Francisco, CA, June 1995.
[39] P. R. Cohen and C. R. Perrault. Elements of a plan based theory of speech acts.
Cognitive Science, 3:177-212, 1979.

[40] D. Connah and P. Wavish. An experiment in cooperation. In Y. Demazeau and J. ­

P. MUller, editors,
Decentralized AI - Proceedings of the First European Work­
shop on Modelling Autonomous Agents in a Multi-Agent World (MAAMAW-89),
pages 197-214. Elsevier Science Publishers B. Y.: Amsterdam, The Netherlands,
1990.

[41] N. 1. Davies. A first-order theory of truth, knowledge and belief. In

Logics in AI
- Proceedings of the European Workshop JELIA-90 (LNAI Volume 478), pages
170-179. Springer-Verlag: Berlin, Germany, 1991.

[42] D. C. Dennett. Brainstorms. The MIT Press: Cambridge, MA 1978.,

[43] D. C. Dennett. The Intentional Stance. The MIT Press: Cambridge, MA 1987. ,

[44] D. C. Dennett. Kinds of Minds. London: Phoenix, 1996.

[45] J. des Rivieres and H. J. Levesque. The consistency of syntactical treatments

of knowledge. In 1. Y. Halpern, editor,
Proceedings of the 1986 Coriference on
Theoretical Aspects of Reasoning About Knowledge, pages 115-l30. Morgan
Kaufmann Publishers: San Mateo, CA, 1986.
208 References

[46] M. d'Inverno, D. Kinny, M. Luck, and M. Wooldridge. A formal specification

of dMARS. In M. P. Singh, A. Rao, and M. J. Wooldridge, editors, Intelligent
Agents I V (LNAI Volume 1365), pages 155-176. Springer-Verlag: Berlin, Ger­
many, 1997.

[47] C. Dixon, M. Fisher, and M. Wooldridge. Resolution for temporal logics of

knowledge. Journal of Logic and Computation, 8(3):345-372, 1998.

[48] C. Eliasmith. Dictionary of the philosophy of mind. Online at

http://www. artsci. wustl. edurphilos/MindDict/, 1999.

[49] E. A. Emerson. Temporal and modal logic. In J. van Leeuwen, editor, Hand­
book of Theoretical Computer Science Volume B: Formal Models and Semantics,
pages 996-1072. Elsevier Science Publishers B. Y.: Amsterdam, The Nether­
lands, 1990.

[50] E. A. Emerson and J. Y. Halpern. 'Sometimes' and 'not never' revisited:

on branching time versus linear time temporal logic. Journal of the ACM,
33(1):151-178, 1986.

[51] E. A. Emerson and 1. Srinivasan. Branching time logic. In J. W. de Bakker,

w. -P. de Roever, and G. Rozenberg, editors, REX School-Workshop on Linear
Time, Branching Time and Parial Order in Logics and Models for Concurrency
(LNCS Volume 354), pages 123-172. Springer-Verlag: Berlin, Germany, 1988.
[52] H. B. Enderton. A Mathematical Introduction to Logic. The Academic Press:
London, England, 1972.

[53] O. Etzioni. Intelligence without robots. AI Magazine, 14(4), December 1993.

[54] R. Fagin, 1. Y. Halpern, Y. Moses, and M. Y. Vardi. Reasoning About Knowledge.

The MIT Press: Cambridge, MA 1995. ,

[55] R. Fagin, J. Y. Halpern, Y. Moses, and M. Y. Vardi. Knowledge-based programs.

Distributed Computing, 10(4):199-225, 1997.

[56] J. Ferber. Multi-Agent Systems. Addison-Wesley: Reading, MA 1999. ,

[57] I. A. Ferguson. Towards an architecture for adaptive, rational, mobile agents.

In E. Werner and Y. Demazeau, editors,
Decentralized AI 3 Proceedings of
-

the Third European Workshop on Modelling Autonomous Agents in a Multi­

Agent World (MAAMAW-9J), pages 249-262. Elsevier Science Publishers B. Y.:
Amsterdam, The Netherlands, 1992.

[58] I. A. Ferguson. Integrated control and coordinated behaviour: A case for agent
models. In M. Wooldridge and N. R. Jennings, editors, Intelligent Agents:
Theories, Architectures, and Languages (LNAI Volume 890), pages 203-218.
Springer-Verlag: Berlin, Germany, January 1995.

[59] R. E. Fikes and N. Nilsson. STRIPS: A new approach to the application of

theorem proving to problem solving. Artificial Intelligence, 5(2):189-208, 1971.

[60] FIPA. Specification part 2 - Agent communication language, 1999. The text
refers to the specification dated 16 April 1999.
References 209

[61] K. Fischer, J. P. Miiller, and M. Pischel. A pragmatic BDI architecture. In

M. Wooldridge, J. P. Miiller, and M. Tambe, editors, Intelligent Agents II (LNAI
Volume 1037), pages 203-218. Springer-Verlag: Berlin, Germany, 1996.

[62] M. Fisher. A resolution method for temporal logic. In Proceedings of the Twelfth
International Joint Coriference on Artificial Intelligence (IJCAI-91), Sydney,
Australia, August 1991.

[63] M. Fisher. A survey of Concurrent METATEM - the language and its applica­
tions. In D. M. Gabbay and H. 1. Ohlbach, editors, Temporal Logic - Proceed­
ings of the First International Conference (LNAI Volume 827), pages 480-505.
Springer-Verlag: Berlin, Germany, July 1994.

[64] M. Fisher. An introduction to executable temporal logic. The Knowledge

Engineering Review, 11(1):43-56, 1996.

[65] M. Fisher and M. Wooldridge. Executable temporal logic for distributed AI.
In Proceedings of the Twelfth International Workshop on Distributed Artificial
Intelligence (I WDAI-93), pages 131-142, Hidden Valley, PA, May 1993.
[66] M. Fitting. Proof Methods for Modal and Intuitionistic Logics. D. Reidel
Publishing Company: Dordrecht, The Netherlands, 1983. (Synthese library
volume 169).

[67] The Foundation for Intelligent Physical Agents. See http://www. fipa. org/.

[68] J. R. Galliers. A Theoretical Framework for Computer Models of Cooperative

Dialogue, Acknowledging Multi-Agent Conflict. PhD thesis, Open University,
UK, 1988.

[69] 1. R. Galliers. Cooperative interaction as strategic belief revision. In S. M.

Deen, editor,
CKBS-90 - Proceedings of the International Working Conference
on Cooperating Knowledge Based Systems, pages 148-163. Springer-Verlag:
Berlin, Germany, 1991.

[70] A Galton. Temporal logic and computer science: An overview. In A Galton,

editor, Temporal Logics and their Applications, pages 1-52. The Academic
Press: London, England, 1987.

[71] P. Giirdenfors. Knowledge in Flux. The MIT Press: Cambridge, MA, 1988.

[72] C. Geissler and K. Konolige. A resolution method for quantified modal logics
of knowledge and belief. In J. Y. Halpern, editor,
Proceedings of the 1986
Conference on Theoretical Aspects of Reasoning About Knowledge, pages 309-
324. Morgan Kaufmann Publishers: San Mateo, CA, 1986.

[73] M. R. Genesereth and N. Nilsson. Logical Foundations of Artificial Intelligence.

Morgan Kaufmann Publishers: San Mateo, CA, 1987.

[74] M. Georgeff, B. Pell, M. Pollack, M. Tambe, and M. Wooldridge. The belief­

desire-intention model of agency. In 1. P. Miiller, M. P. Singh, and A S. Rao,
editors, Intelligent Agents V (LNAI Volume 1555), pages 1-10. Springer-Verlag:
Berlin, Germany, 1999.
210 References

[75] M. P. Georgeff. Communication and interaction in multi-agent planning. In

Proceedings of the Third National Conforence on Artificial Intelligence (AAAI-
83), Washington, D. C. , 1983.
[76] M. P. Georgeff. A theory of action for multi-agent planning. In Proceedings
of the Fourth National Conference on Artificial Intelligence (AAAI-84), Austin,
TX, 1984.

[77] M. P. Georgeff and F. F. Ingrand. Decision-making in an embedded reasoning

system. In Proceedings of the Eleventh International Joint Conference on
Artificial Intelligence (IJCAI-89), pages 972-978, Detroit, MI, 1989.
[78] M. P. Georgeff and A. L. Lansky. Reactive reasoning and planning. In Proceed­
ings of the Sixth National Conforence on Artificial Intelligence (AAAI-87), pages
677-682, Seattle, WA, 1987.

[79] M. P. Georgeff and A. S. Rao. The semantics of intention maintenance for

rational agents. In Proceedings of the Fourteenth International Joint Conference
on Artificial Intelligence (IJCAI-95), pages 704-710, Montreal, Quebec, Canada,
August 1995.

[80] M. P. Georgeff and A. S. Rao. A profile of the Australian AI Institute. IEEE

Expert, 11(6):89-92, December 1996.

[81] R. Goldblatt. Logics of Time and Computation (CSLI Lecture Notes Number 7).
Center for the Study of Language and Information, Ventura Hall, Stanford, CA
94305, 1987. (Distributed by Chicago University Press).

[82] G. D. Gough. Decision procedures for temporal logic. Master's thesis, De­
partment of Computer Science, Manchester University, Oxford Rd. , Manchester
M13 9PL, UK, October 1984.

[83] M. Greaves and 1. Bradshaw, editors. SpecifYing and Implementing Conversation

Policies - Proceedings of the Autonomous Agents 2000 Workshop. Seattle, WA,
May 1999.

[84] B. 1. Grosz and S. Kraus. Collaborative plans for complex group action. Artificial
Intelligence, 86(2):269-357, 1996.

[85] B. 1. Grosz and S. Kraus. The evolution of SharedPlans. In M. Wooldridge

and A. Rao, editors, Foundations of Rational Agency, pages 227-262. Kluwer
Academic Publishers: Boston, MA, 1999.

[86] A. Haddadi. Towards a pragmatic theory of interactions. In Proceedings of

the First International Conference on Multi-Agent Systems (ICMAS-95), pages
133-139, San Francisco, CA, June 1995.

[87] A. Haddadi. Communication and Cooperation in Agent Systems (LNAI Volume

1056). Springer-Verlag: Berlin, Germany, 1996.

[88] J. Y. Halpern. Using reasoning about knowledge to analyze distributed systems.

Annual Review of Computer Science, 2:37-68, 1987.
References 211

[89] J. Y. Halpern. Knowledge and common knowledge in a distributed environment.

Journal of the ACM, 37(3), 1990.

[90] J. Y. Halpern and Y. Moses. A guide to completeness and complexity for modal
logics of knowledge and belief. Artificial Intelligence, 54:319-379, 1992.

[91] J. Y. Halpern and M. Y. Vardi. The complexity of reasoning about knowledge and
time. I. Lower bounds. Journal of Computer and System Sciences, 38:195-237,
1989.

[92] J. Y. Halpern and M. Y. Vardi. Model checking versus theorem proving: A

manifesto. In V. Lifschitz, editor,
AI and Mathematical Theory of Computation
- Papers in Honor of John McCarthy, pages 151-176. The Academic Press:
London, England, 1991.

[93] D. Harel. Dynamic logic. In D. Gabbay and F. Guenther, editors, Handbook of

Philosophical Logic Volume II - Extensions of Classical Logic, pages 497--604.
D. Reidel Publishing Company: Dordrecht, The Netherlands, 1984. (Synthese
library Volume 164).

[94] K. V. Hindriks, F. S. de Boer, W. van der Hoek, and J. -J. Ch. Meyer. Formal
semantics for an abstract agent programming language. In M. P. Singh, A. Rao,
and M. J. Wooldridge, editors, Intelligent Agents I V (LNAI Volume 1365), pages
215-230. Springer-Verlag: Berlin, Germany, 1998.

[95] K. V. Hindriks, F. S. de Boer, W. van der Hoek, and J. -J. Ch. Meyer. Control
structures of rule-based agent languages. In J. P. MUller, M. P. Singh, and A. S.
Rao, editors, Intelligent Agents V (LNAI Volume 1555). Springer-Verlag: Berlin,
Germany, 1999.

[96] J. Hintikka. Knowledge and Belief Cornell University Press: Ithaca, NY, 1962.

[97] C. A. R. Hoare. An axiomatic basis for computer programming. Communica­

tions of the ACM, 12(10):576-583, 1969.

[98] C. A. R. Hoare. Communicating sequential processes. Communications of the

ACM, 21:666--677, 1978.

[99] W. van der Hoek, B. van Linder, and J. -J. Ch. Meyer. An integrated modal
approach to rational agents. In M. Wooldridge and A. Rao, editors, Foundations
of Rational Agency, pages 133-168. Kluwer Academic Publishers: Boston, MA,
1999.

[100] M. Huber. JAM: A BDI-theoretic mobile agent architecture. In Proceedings of

the Third International Conference on Autonomous Agents (Agents 99), pages
236-243, Seattle, WA, 1999.

[101] G. E. Hughes and M. J. Cresswell. Introduction to Modal Logic. Methuen and

Co. , Ltd. , 1968.

[102] M. Huhns and M. P. Singh, editors. Readings in Agents. Morgan Kaufmann

Publishers: San Mateo, CA, 1998.
212 References

[103] D. 1. Israel. The role(s) of logic in artificial intelligence. In D. M. Gabbay,

C. 1. Hogger, and 1. A. Robinson, editors, Handbook of Logic in Artificial
Intelligence and Logic Programming, pages 1-29. Oxford University Press:
Oxford, England, 1993.

[104] N. R. Jennings. On being responsible. In E. Werner and Y. Demazeau, ed­

itors,
Decentralized AI 3 - Proceedings of the Third European Workshop on
Modelling Autonomous Agents in a Multi-Agent World (MAAMAW-91), pages
93-102. Elsevier Science Publishers B. Y.: Amsterdam, The Netherlands, 1992.

[105] N. R. Jennings. Commitments and conventions: The foundation of coordination

in multi-agent systems. The Knowledge Engineering Review, 8(3):223-250,
1993.

[106] N. R. Jennings. Controlling cooperative problem solving in industrial multi­

agent systems using joint intentions. Artificial Intelligence, 74(2), 1995.

[107] N. R. Jennings. Agent-based computing: Promise and perils. In Proceedings

of the Sixteenth International Joint Conference on Artificial Intelligence (IJCAI-
99), pages 1429-1436, Stockholm, Sweden, 1999.
[108] N. R. Jennings, K. Sycara, and M. Wooldridge. A roadmap of agent research and
development. Autonomous Agents and Multi-Agent Systems, 1(1):7-38, 1998.

[109] N. R. Jennings and M. Wooldridge, editors. Agent Technology: Foundations,

Applications and Markets. Springer-Verlag: Berlin, Germany, 1998.

[110] C. B. Jones. Systematic Software Development using VDM (second edition).

Prentice Hall, 1990.

[111] C. G. Jung. Emergent mental attitudes in layered agents. In 1. P. Muller,

M. P. Singh, and A. S. Rao, editors, Intelligent Agents V (LNAI Volume 1555).
Springer-Verlag: Berlin, Germany, 1999.

[112] L. P. Kaelbling. An architecture for intelligent reactive systems. In M. P.

Georgeff and A. L. Lansky, editors, Reasoning About Actions & Plans - Pro­
ceedings of the 1986 Workshop, pages 395-410. Morgan Kaufmann Publishers:
San Mateo, CA, 1986.

[113] L. P. Kaelbling. Learning in Embedded Systems. The MIT Press: Cambridge,

MA, 1993.

[114] L. P. Kaelbling and S. J. Rosenschein. Action and planning in embedded agents.

In P. Maes, editor, Designing Autonomous Agents, pages 35-48. The MIT Press:
Cambridge, MA, 1990.

[115] J. A. W. Kamp. Tense Logic and the Theory of Linear Order. PhD thesis,
University of California, 1968.

[116] D. Kinny and M. Georgeff. Commitment and effectiveness of situated agents.

In Proceedings of the Twelfth International Joint Conference on Artificial Intel­
ligence (IJCAI-91), pages 82-88, Sydney, Australia, 1991.
References 213

[117] D. Kinny, M. Georgeft', and A. Rao. A methodology and modelling technique

for systems of BDI agents. In W. Van de Ve1de and J. W. Perram, editors,
Agents
Breaking Away: Proceedings of the Seventh European Workshop on Modelling
Autonomous Agents in a Multi-Agent World, (LNAI Volume 1038), pages 56-71.
Springer-Verlag: Berlin, Germany, 1996.

[118] D. Kinny, M. Ljungberg, A. S. Rao, E. Sonenberg, G. Tidhar, and E. Werner.

Planned team activity. In C. Castelfranchi and E. Werner, editors, Artificial
Social Systems - Selected Papers from the Fourth European Workshop on
Modelling Autonomous Agents in a Multi-Agent World, MAAMAW-92 (LNAI
Volume 830), pages 226-256. Springer-Verlag: Berlin, Germany, 1992.
[119] K. Konolige. A first-order formalization of knowledge and action for a multi­
agent planning system. In 1. E. Hayes, D. Michie, and Y. Pao, editors, Machine
Intelligence 10, pages 41-72. Ellis Horwood: Chichester, England, 1982.

[120] K. Konolige. A Deduction Model of Belief Pitman Publishing: London and

Morgan Kaufmann: San Mateo, CA, 1986.

[121] K. Konolige. Hierarchic autoepistemic theories for nonmonotonic reasoning:

Preliminary report. In M. Reinfrank, 1. de Kleer, M. L. Ginsberg, and E. Sande­
wall, editors, Nonmonotonic Reasoning - Proceedings of the Second Interna­
tional Workshop (LNAI Volume 346), pages 42-59. Springer-Verlag: Berlin, Ger­
many, 1988.

[122] D. Kozen and J. Tiuryn. Logics of programs. In 1. van Leeuwen, editor,

Handbook of Theoretical Computer Science Volume B: Formal Models and
Semantics, pages 789-840. Elsevier Science Publishers B. Y.: Amsterdam, The
Netherlands, 1990.

[123] S. Kraus. Negotiation and cooperation in multi-agent environments. Artificial

Intelligence, 94(1-2):79-98, July 1997.

[124] S. Kraus, M. Nirke, and K. Sycara. Reaching agreements through argumentation:

A logical model. In Proceedings of the Twelfth International Workshop on
Distributed Artificial Intelligence (I WDAI-93), pages 233-247, Hidden Valley,
PA, May 1993.

[125] S. Kraus, K. Sycara, and A. Evenchik. Reaching agreements through argumen­

tation: a logical model and implementation. Artificial Intelligence, 104:1--69,
1998.

[126] S. Kripke. Semantical analysis of modal logic. Zeitschrift for Mathematische

Logik und Grundlagen der Mathematik, 9:67-96, 1963.

[127] Y. Labrou and T. Finin. Semantics and conversations for an agent communica­
tion language. In Proceedings of the Fifteenth International Joint Conference on
Artificial Intelligence (IJCAI-97), pages 584-591, Nagoya, Japan, 1997.
[128] Y. Labrou and T. Finin. Agent communication languages: The current landscape.
IEEE Intelligent Systems, 14(2):45-52, 1999.
214 References

[129] L. Lamport. Sometimes is sometimes not never - but not always. In Proceed­
ings of the Seventh ACM Symposium on the Principles of Programming Lan­
guages (POPL), 1980.
[130] Y. Lesperance, H. J. Levesque, F. Lin, D. Marcu, R. Reiter, and R. B. Scher!.
Foundations of a logical approach to agent programming. In M. Wooldridge,
J. P. Miiller, and M. Tambe, editors, Intelligent Agents II (LNAI Volume 1037),
pages 331-346. Springer-Verlag: Berlin, Germany, 1996.

[131] H. Levesque, R. Reiter, Y. Lesperance, F. Lin, and R. Scher!. Golog: A logic

programming language for dynamic domains. Journal of Logic Programming,
31:59-84, 1996.

[132] H. 1. Levesque. A logic of implicit and explicit belief. In Proceedings of the

Fourth National Conference on Artificial Intelligence (AAAI-84), pages 198-
202, Austin, TX, 1984.

[133] H. 1. Levesque, P. R. Cohen, and 1. H. T. Nunes. On acting together. In

Proceedings of the Eighth National Conforence on Artificial Intelligence (AAAI-
90), pages 94-99, Boston, MA, 1990.
[134] o. Lichtenstein and A. Pnueli. Checking that finite state concurrent programs
satisfY their linear specification. In Proceedings of the Eleventh ACM Symposium
on the Principles of Programming Languages, pages 97-107, 1984.

[135] O. Lichtenstein and A. Pnueli. Propositional temporal logics: Decidability and

completeness. Logic Journal of the IGPL, 8(1):55-85, 2000.

[136] V. Lifschitz. On the semantics of STRIPS. In M. P. Georgeff and A. L. Lansky,

editors, Reasoning About Actions & Plans - Proceedings of the 1986 Workshop,
pages 1-10. Morgan Kaufmann Publishers: San Mateo, CA, 1986.

[137] B. van Linder, W. van der Hoek, and 1. -J. Ch. Meyer. How to motivate your
agents. In M. Wooldridge, J. P. Miiller, and M. Tambe, editors, Intelligent Agents
II (LNAI Volume 1037), pages 17-32. Springer-Verlag: Berlin, Germany, 1996.

[138] M. Ljunberg and A. Lucas. The OASIS air traffic management system. In
Proceedings of the Second Pacific Rim International Conference on AI (PRICAI-
92), Seoul, Korea, 1992.
[139] A. Lomuscio. Knowledge Sharing among Ideal Agents. PhD thesis, School of
Computer Science, University of Birmingham, Birmingham, UK, June 1999.

[140] A. Lomuscio and M. Ryan. A spectrum of modes of knowledge sharing be­

tween agents. In N.R. Jennings and Y. Lesperance, editors,
Intelligent Agents
VI - Proceedings of the Sixth International Workshop on Agent Theories, Ar­
chitectures, and Languages (ATAL-99), Lecture Notes in Artificial Intelligence.
Springer-Verlag, Berlin, 2000.

[141] M. Luck, N. Griffiths, and M. d'Invemo. From agent theory to agent construc­
tion: A case study. In 1. P. Miiller, M. Wooldridge, and N. R. Jennings, editors,
Intelligent Agents III (LNAI Volume 1193), pages 49-64. Springer-Verlag: Berlin,
Germany, 1997.
References 215

[142] P. Maes. Agents that reduce work and information overload. Communications
of the ACM, 37(7):31-40, July 1994.

[143] Z. Manna and A. Pnueli. Verification of concurrent programs: The temporal

framework. In R. S. Boyer and I. S. Moore, editors, The Correctness Problem
in Computer Science, pages 215-273. The Academic Press: London, England,
1981.

[144] Z. Manna and A. Pnueli. The Temporal Logic of Reactive and Concurrent
Systems. Springer-Verlag: Berlin, Germany, 1992.

[145] Z. Manna and A. Pnueli. Temporal Verification of Reactive Systems - Safety.

Springer-Verlag: Berlin, Germany, 1995.

[146] Z. Manna and P. Wolper. Synthesis of communicating processes from temporal

logic specifications. ACM Transactions on Programming Languages and Sys­
tems, 6(1):68-93, January 1984.

[147] J. Mayfield, Y. Labrou, and T. Finin. Evaluating KQML as an agent communica­

tion language. In M. Wooldridge, I. P. Miiller, and M. Tambe, editors, Intelligent
Agents II (LNAI Volume 1037), pages 347-360. Springer-Verlag: Berlin, Ger­
many, 1996.

[148] J. McCarthy. Ascribing mental qualities to machines. Technical report, Stanford

University AI Lab. , Stanford, CA 94305, 1978.

[149] J. McCarthy. Formalization of common sense: papers by John McCarthy. Ablex

Publishing Corp. , 1990.

[150] J. McCarthy. Modality, Si! Modal logic, No! Studia Logica, 59:29-32, 1997.

[151] J. McCarthy and P. I. Hayes. Some philosophical problems from the standpoint
of artificial intelligence. In B. Meltzer and D. Michie, editors, Machine Intelli­
gence 4. Edinburgh University Press, 1969.

[152] J. -I. Ch. Meyer and W. van der Hoek. Epistemic Logic for AI and Computer
Science. Cambridge University Press: Cambridge, England, 1995.

[153] J.-I. Ch. Meyer, W. van der Hoek, and B. van Linder. A logical approach to the
dynamics of commitments. Artificial Intelligence, 113:1-40, 1999.

[154] R. Milner. Communication and Concurrency. Prentice Hall, 1989.

[155] R. Montague. Syntactical treatments of modality, with corollaries on reflexion

principles and finite axiomatizations. Acta Philosophica Fennica, 16:153-167,
1963.

[156] R. C. Moore. Reasoning about knowledge and action. In Proceedings of the Fifth
International Joint Conference on Artificial Intelligence (/JCAI-77), Cambridge,
MA, 1977.

[157] R. C. Moore. A formal theory of knowledge and action. In I. F. Allen, I. Hendler,

and A. Tate, editors, Readings in Planning, pages 480-519. Morgan Kaufmann
Publishers: San Mateo, CA, 1990.
216 References

[158] M. Mora, 1. G. Lopes, R. Viccari, and H. Coelho. BD! models and systems:
reducing the gap. In J. P. MUller, M. P. Singh, and A. S. Rao, editors, Intelligent
Agents V (LNAI Volume 1555). Springer-Verlag: Berlin, Germany, 1999.

[159] C. Morgan. Programming from Specifications (second edition). Prentice Hall

International: Hemel Hempstead, England, 1994.

[160] L. Morgenstern. A first-order theory of planning, knowledge, and action. In J. Y.

Halpern, editor, Proceedings of the 1986 Conference on Theoretical Aspects of
Reasoning About Knowledge, pages 99-114. Morgan Kaufmann Publishers: San
Mateo, CA, 1986.

[161] L. Morgenstern. Knowledge preconditions for actions and plans. In Proceedings

of the Tenth International Joint Conference on Artificial Intelligence (IJCAI-87),
pages 867-874, Milan, Italy, 1987.

[162] J. P. MUller. The Design of Intelligent Agents (LNAI Volume 1177). Springer­
Verlag: Berlin, Germany, 1997.

[163] J. P. MUller, M. Pischel, and M. Thiel. Modelling reactive behaviour in vertically

layered agent architectures. In M. Wooldridge and N. R. Jennings, editors,
Intelligent Agents: Theories, Architectures, and Languages (LNAI Volume 890),
pages 261-276. Springer-Verlag: Berlin, Germany, January 1995.

[164] J. P. MUller, M. Wooldridge, and N. R. Jennings, editors. Intelligent Agents III

(LNAI Volume 1193). Springer-Verlag: Berlin, Germany, 1997.

[165] N. Muscettola, P. Pandurang Nayak, B. Pell, and B. C. Williams. Remote agents:

To boldly go where no AI system has gone before. Artificial Intelligence, 103:5-
47, 1998.

[166] A. Newell. The knowledge level. Artificial Intelligence, 18(1):87-127, 1982.

[167] N. 1. Nilsson. Towards agent programs with circuit semantics. Technical Re­
port STAN-CS-92-1412, Computer Science Department, Stanford University,
Stanford, CA 94305, January 1992.

[168] P. Noriega and C. Sierra, editors. Agent Mediated Electronic Commerce (LNAI
Volume 1571). Springer-Verlag: Berlin, Germany, 1999.

[169] S. Parsons, C. A. Sierra, and N. R. Jennings. Agents that reason and negotiate
by arguing. Journal of Logic and Computation, 8(3):261-292, 1998.

[170] R. S. Patil, R. E. Fikes, P. F. Patel-Schneider, D. McKay, T. Finin, T. Gruber, and

R. Neches. The DARPA knowledge sharing effort: Progress report. In C. Rich,
W. Swartout, and B. Nebel, editors, Proceedings of Knowledge Representation
and Reasoning (KR&R-92), pages 777-788, 1992.

[171] D. Perlis. Languages with self reference I: Foundations. Artificial Intelligence,

25:301-322, 1985.

[172] D. Perlis. Languages with self reference II: Knowledge, belief, and modality.
Artificial Intelligence, 34:179-212, 1988.
References 217

[173] D. Perlis. Meta in logic. In P. Maes and D. Nardi, editors, Meta-Level Architec­
tures and Reflection, pages 37-49. Elsevier Science Publishers B. Y.: Amsterdam,
The Netherlands, 1988.

[174] C. R. Perrault. An application of default logic to speech acts theory. In P. R.

Cohen, J. Morgan, and M. E. Pollack, editors, Intentions in Communication,
pages 161-186. The MIT Press: Cambridge, MA 1990. ,

[175] J. Pitt and E. H. Mamdani. A protocol-based semantics for an agent communi­

cation language. In Proceedings of the Sixteenth International Joint Coriference
on Artificial Intelligence (/JCAI-99), Stockholm, Sweden, August 1999.
[176] A. Pnueli. The temporal logic of programs. In Proceedings of the Eighteenth
IEEE Symposium on the Foundations of Computer Science, pages 46-57, 1977.
[177] A. Pnueli. Specification and development of reactive systems. In Information
Processing 86. Elsevier Science Publishers B. Y.: Amsterdam, The Netherlands,
1986.

[178] A. Pnueli and R. Rosner. On the synthesis of a reactive module. In Proceedings

of the Sixteenth ACM Symposium on the Principles of Programming Languages
(POPL), pages 179-190, January 1989.
[179] M. E. Pollack. Plans as complex mental attitudes. In P. R. Cohen, 1. Morgan,
and M. E. Pollack, editors, Intentions in Communication, pages 77-104. The
MIT Press: Cambridge, MA 1990. ,

[180] M. E. Pollack. The uses of plans. Artificial Intelligence, 57(1):43-68, September

1992.

[181] R. Power. Mutual intention. Journal for the Theory of Social Behaviour, 14:85-
102, March 1984.

[182] A. Prior. Past, Present and Future. Oxford University Press: Oxford, England,
1967.

[183] A. S. Rao. AgentSpeak(L): BDI agents speak out in a logical computable lan­
guage. In W. Van de Velde and 1. W. Perram, editors, Agents Breaking Away:
Proceedings of the Seventh European Workshop on Modelling Autonomous
Agents in a Multi-Agent World, (LNAI Volume 1038), pages 42-55. Springer­
Verlag: Berlin, Germany, 1996.

[184] A. S. Rao. Decision procedures for propositional linear-time Belief-Desire­

Intention logics. In M. Wooldridge, J. P. MUller, and M. Tambe, editors, In­
telligent Agents II (LNAI Volume 1037), pages 33-48. Springer-Verlag: Berlin,
Germany, 1996.

[185] A. S. Rao and M. Georgeff. Decision procedures for BDI logics. Journal of
Logic and Computation, 8(3):293-344, 1998.

[186] A. s. Rao and M. P. Georgeff. Asymmetry thesis and side-effect problems in

linear time and branching time intention logics. In Proceedings of the Twelfth
International Joint Conference on Artificial Intelligence (/JCAI-91), pages 498-
504, Sydney, Australia, 1991.
218 References

[187] A. S. Rao and M. P. Georgeft'. Modeling rational agents within a BDI­

architecture. In R. Fikes and E. Sandewall, editors, Proceedings of Knowledge
Representation and Reasoning (KR&R-91), pages 473-484. Morgan Kaufmann
Publishers: San Mateo, CA, April 1991.

[188] A. S. Rao and M. P. Georgeft'. An abstract architecture for rational agents.

In C. Rich, W. Swartout, and B. Nebel, editors, Proceedings of Knowledge
Representation and Reasoning (KR&R-92), pages 439-449, 1992.

[189] A. S. Rao and M. P. Georgeft'. A model-theoretic approach to the verification of

situated reasoning systems. In Proceedings of the Thirteenth International Joint
Coriference on Artificial Intelligence (IJCAI-93), pages 318-324, Chambery,
France, 1993.

[190] A. S. Rao, M. P. Georgeft', and E. A. Sonenberg. Social plans: A preliminary

report. In E. Werner and Y. Demazeau, editors,
Decentralized AI 3 - Proceed­
ings of the Third European Workshop on Modelling Autonomous Agents in a
Multi-Agent World (MAAMAW-91), pages 57-76. Elsevier Science Publishers
B. Y.: Amsterdam, The Netherlands, 1992.

[191] H. Reichgelt. A comparison of first-order and modal logics of time. In P. Jack­

son, H. Reichgelt, and F. van Harmelen, editors, Logic Based Knowledge Repre­
sentation, pages 143-176. The MIT Press: Cambridge, MA 1989. ,

[192] H. Reichgelt. Logics for reasoning about knowledge and belief. The Knowledge
Engineering Review, 4(2):119-139, 1989.

[193] R. Reiter. A logic for default reasoning. Artificial Intelligence, 13:81-132, 1980.

[194] J. S. Rosenschein. Synchronisation of multi-agent plans. In Proceedings of the

Second National Conforence on Artificial Intelligence (AAAI-82), Pittsburgh, PA,
1982.

[195] J. S. Rosenschein and G. Zlotkin. Rules of Encounter: Designing Conventions

for Automated Negotiation among Computers. The MIT Press: Cambridge, MA ,

1994.

[196] S. Rosenschein and L. P. Kaelbling. The synthesis of digital machines with

provable epistemic properties. In J. Y. Halpern, editor,
Proceedings of the 1986
Coriference on Theoretical Aspects of Reasoning About Knowledge, pages 83-
98. Morgan Kaufmann Publishers: San Mateo, CA, 1986.

[197] S. J. Rosenschein and L. P. Kaelbling. A situated view of representation and

control. In P. E. Agre and S. J. Rosenschein, editors, Computational Theories of
Interaction and Agency, pages 515-540. The MIT Press: Cambridge, MA 1996. ,

[198] S. Russell and P. Norvig. Artificial Intelligence: A Modern Approach. Prentice­

Hall, 1995.

[199] T. Sandbolm. Distributed rational decision making. In G. WeiB, editor, Multia­

gent Systems, pages 201-258. The MIT Press: Cambridge, MA 1999. ,
References 219

[200] K. Schild. On the relationship between BD! logics and standard logics of
concurrency. In 1. P. Muller, M. P. Singh, and A. S. Rao, editors, Intelligent
Agents V (LNAI Volume 1555). Springer-Verlag: Berlin, Germany, 1999.

[201] M. 1. Schoppers. Universal plans for reactive robots in unpredictable environ­

ments. In Proceedings of the Tenth International Joint Conference on Artificial
Intelligence (/JCAI-B7), pages 1039-1046, Milan, Italy, 1987.
[202] J. R. Searle. Speech Acts: An Essay in the Philosophy of Language. Cambridge
University Press: Cambridge, England, 1969.

[203] J. R. Searle. Collective intentions and actions. In P. R. Cohen, 1. Morgan, and

M. E. Pollack, editors, Intentions in Communication, pages 401-416. The MIT
Press: Cambridge, MA 1990.,

[204] N. See!. Agent Theories and Architectures. PhD thesis, Surrey University,
Guildford, UK, 1989.

[205] Y. Shoham. Agent-oriented programming. Technical Report STAN-CS-1335-

90, Computer Science Department, Stanford University, Stanford, CA 94305,
1990.

[206] Y. Shoham. Agent-oriented programming. Artificial Intelligence, 60(1):51-92,

1993.

[207] Carles Sierra, Nick R. Jennings, Pablo Noriega, and Simon Parsons. A frame­
work for argumentation-based negotiation. In M. P. Singh, A. Rao, and M. J.
Wooldridge, editors, Intelligent Agents IV (LNAI Volume 1365), pages 177-192.
Springer-Verlag: Berlin, Germany, 1998.

[208] H. A. Simon. The Sciences of the Artificial (second edition). The MIT Press:
Cambridge, MA 1981.
,

[209] M. Singh. Agent communication languages: Rethinking the principles. IEEE

Computer, pages 40-47, December 1998.

[210] M. P. Singh. Group intentions. In Proceedings of the Tenth International

Workshop on Distributed Artificial Intelligence (I WDAI-90), 1990.
[211] M. P. Singh. Towards a theory of situated know-how. In Proceedings of the
Ninth European Conference on Artificial Intelligence (ECAI-90), pages 604-
609, Stockholm, Sweden, 1990.

[212] M. P. Singh. Group ability and structure. In Y. Demazeau and J. -P. Muller,
editors,
Decentralized AI 2 Proceedings of the Second European Workshop
-

on Modelling Autonomous Agents in a Multi-Agent World (MAAMAW-90), pages

127-146. Elsevier Science Publishers B. V. : Amsterdam, The Netherlands, 1991.

[213] M. P. Singh. Social and psychological commitments in multiagent systems. In

AAAI Fall Symposium on Knowledge and Action at Social and Organizational
Levels, 1991.
[214] M. P. Singh. Towards a formal theory of communication for multi-agent sys­
tems. In Proceedings of the Twelfth International Joint Conference on Artificial
Intelligence (/JCAI-91), pages 69-74, Sydney, Australia, 1991.
220 References

[215] M. P. Singh. A semantics for speech acts. Annals of Mathematics and Artificial
Intelligence, 8(1-11):47-71, 1993.

[216] M. P. Singh. Multiagent Systems: A Theoretical Framework for Intentions,

Know-How, and Communications (LNAI Volume 799). Springer-Verlag: Berlin,
Germany, 1994.

[217] M. P. Singh. The intentions of teams: Team structure, endodeixis, and exodeixis.
In Proceedings of the Thirteenth European Conference on Artificial Intelligence
(ECAI-98), pages 303-307, Brighton, United Kingdom, 1998.
[218] M. P. Singh and N. M. Asher. Towards a formal theory of intentions. In J. van
Eijck, editor,
Logics in AI - Proceedings of the European Workshop JELIA-90
(LNAI Volume 478), pages 472-486. Springer-Verlag: Berlin, Germany, 1991.
[219] A. Sistla. Theoretical issues in the design and verification of distributed systems.
Technical Report CMU-CS-83-146, School of Computer Science, Carnegie­
Mellon University, Pittsburgh, PA, 1983.

[220] A. P. Sistla and E. M. Clarke. The complexity of propositional linear temporal

logics. Journal of the ACM, 32(3):733-749, 1985.

[221] A. P. Sistla, E. M. Clarke, N. Francez, and A. R. Meyer. Can message buffers be

axiomatized in temporal logic? Information and Control, 63(112):88-112, 1985.

[222] R. G. Smith. A Framework for Distributed Problem Solving. UMI Research

Press, 1980.

[223] R. M. Smullyan. First-Order Logic. Springer-Verlag: Berlin, Germany, 1968.

[224] M. Spivey. The Z Notation (second edition). Prentice Hall International: Hemel
Hempstead, England, 1992.

[225] L. Steels. Cooperation between distributed agents through self organization.

In Y. Demazeau and J. -P. Muller, editors,
Decentralized AI - Proceedings of
the First European Workshop on Modelling Autonomous Agents in a Multi­
Agent World (MAAMAW-89), pages 175-196. Elsevier Science Publishers B. Y.:
Amsterdam, The Netherlands, 1990.

[226] S. P. Stich. From Folk Psychology to Cognitive Science. The MIT Press:
Cambridge, MA, 1983.

[227] C. Stirling. Completeness results for full branching time logic. In REX School­
Workshop on Linear Time, Branching Time, and Partial Order in Logics and
Models for Concurrency, Noordwijkerhout, Netherlands, 1988.
[228] K. P. Sycara. Multiagent compromise via negotiation. In L. Gasser and
M. Huhns, editors, Distributed Artificial Intelligence Volume II, pages 119-138.
Pitman Publishing: London and Morgan Kaufmann: San Mateo, CA, 1989.

[229] R. Thomason. A note on syntactical treatments of modality. Synthese, 44:391-

395, 1980.

[230] R. Tuomela and K. Miller. We-intentions. Philosophical Studies, 53:367-389,

1988.
References 221

[231] R. Turner. Truth and Modality for Knowledge Representation. Pitman Publish­
ing: London, 1990.

[232] M. Y. Vardi and P. Wolper. Reasoning about infinite computations. Information

and Computation, 115(1):1-37, 1994.

[233] G. WeiB, editor. Multi-Agent Systems. The MIT Press: Cambridge, MA 1999.
,

[234] E. Werner. W hat can agents do together: A semantics of co-operative ability. In

Proceedings of the Ninth European Conference on Artificial Intelligence (ECAI-
90), pages 694-701, Stockholm, Sweden, 1990.
[235] P. Wolper. Temporal logic can be more expressive. Information and Control, 56,
1983.

[236] P. Wolper. The tableau method for temporal logic: An overview. Logique et
Analyse, 110-111, 1985.

[237] M. Wooldridge. The Logical Modelling of Computational Multi-Agent Systems.

PhD thesis, Department of Computation, UMIST, Manchester, UK, October
1992.

[238] M. Wooldridge. This is MYWORLD: The logic of an agent-oriented testbed for

DAI. In M. Wooldridge and N. R. Jennings, editors, Intelligent Agents: Theories,
Architectures, and Languages (LNAI Volume 890), pages 160-178. Springer­
Verlag: Berlin, Germany, January 1995.

[239] M. Wooldridge. Practical reasoning with procedural knowledge: A logic of

BDI agents with know-how. In D. M. Gabbay and H. -J. Ohlbach, editors,
Practical Reasoning - Proceedings of the International Conference on Formal
and Applied Practical Reasoning, FAPR-96 (LNAI Volume 1085), pages 663-
678. Springer-Verlag: Berlin, Germany, June 1996.

[240] M. Wooldridge. Agent-based software engineering. lEE Proceedings on Soft­

ware Engineering, 144(1):26-37, February 1997.

[241] M. Wooldridge. Verifiable semantics for agent communication languages. In

Proceedings of the Third International Conference on Multi-Agent Systems
(ICMAS-98), pages 349-365, Paris, France, 1998.
[242] M. Wooldridge. Intelligent agents. In G. WeiB, editor, Multiagent Systems, pages
27-78. The MIT Press: Cambridge, MA 1999. ,

[243] M. Wooldridge. Verifying that agents implement a communication language.

InProceedings of the Sixteenth National Conference on Artificial Intelligence
(AAAI-99), pages 52-57, Orlando, FL, July 1999.
[244] M. Wooldridge, C. Dixon, and M. Fisher. A tableau-based proof method for
temporal logics of knowledge and belief. Journal of Applied Non-Classical
Logics, 8(3):225-258, 1998.

[245] M. Wooldridge and M. Fisher. A first-order branching time logic of multi­

agent systems. In
Proceedings of the Tenth European Conference on Artificial
Intelligence (ECAI-92), pages 234-238, Vienna, Austria, 1992.
222 References

[246] M. Wooldridge and M. Fisher. A decision procedure for a temporal belief

logic. In D. M. Gabbay and H. J. Ohlbach, editors, Temporal Logic - Proceed­
ings of the First International Coriference (LNAI Volume 827), pages 317-331.
Springer-Verlag: Berlin, Germany, July 1994.

[247] M. Wooldridge and N. R. Jennings. Formalizing the cooperative problem solv­

ing process. In Proceedings of the Thirteenth International Workshop on Dis­
tributed Artificial Intelligence (I WDAI-94), pages 403-417, Lake Quinalt, WA,
July 1994. Reprinted in [102] .

[248] M. Wooldridge and N. R. Jennings. Towards a theory of cooperative problem

solving. In Proceedings of the Sixth European Workshop on Modelling Au­
tonomous Agents and Multi-Agent Worlds (MAAMAW-94), pages 15-26, August
1994.

[249] M. Wooldridge and N. R. Jennings. Intelligent agents: Theory and practice. The
Knowledge Engineering Review, 10(2):115-152, 1995.

[250] M. Wooldridge and N. R. Jennings. Cooperative problem solving. Journal of

Logic and Computation, 9(4): 563-594, 1999.

[251] M. Wooldridge and S. D. Parsons. Intention reconsideration reconsidered. In

J. P. Miiller, M. P. Singh, and A. S. Rao, editors, Intelligent Agents V (LNAI
Volume 1555), pages 63-80. Springer-Verlag: Berlin, Germany, 1999.
Index

A c

3APL system 177 Castelfranchi, C. 147

3T agent architecture 45 Catach, L. 112
ability 149-154 cautious intention reconsideration 37
accessibility relations 74 choice operator 63
action expressions 62---{j5 choices 56
actions 2, 61--66, 71 Clarke, E. 171
agent control loop 26, 30 Cohen, P. 45, 66, 89, 112, 120, 124, 129, 130,
agent-oriented programming 44 134, 144, 161
agents commissive speech acts 128
embodied 2 commitment 30
rational 1 blind 33
situated 2 open-minded 33
software 18 single-minded 33
Allen, J. 18 strategies 33
always temporal operator 57, 70, 81 composite speech acts 141
answer protocols 141 computation, control of 22, 26
Appelt, D. 144 conative logic 183
arity (of predicates) 48 concurrent actions 67
attempts 134-136 Concurrent METATEM system 168, 173, 176
attitudes, mental 41 conformance testing 144-145
Austin, J. 127 conjunction 48
autonomy 2 constants
axiomatic verification 172-174 interpretation for 76
axiomatizations, complete 91 logical 48, 70
axioms, logical x, 8 contingent truth 185
continuous time 67
Contract Net protocol 154
convention 120
B cooperation 6
cooperative problem solving 147-149
Banieqbal, B. 202 coordinated attack problem 114-115
Barcan formulae 194 correspondence theory 92, 187-189
Barringer, H. 202 for BD! 92
BD! CTL temporal logic 202
correspondence theory 92 CTL * temporal logic 67, 89, 202
logics x, 7
model x, 7-8, 21-40 o
belief
accessibility relation 74 decision theory 10-11, 18
desire consistency 106 declarative speech acts 128
desire weak realism 107 de dicto quantification 55, 196
modality 50 deliberation 21, 22
Ben-Ari, M. 143 Dennett, D. 42
Blair, T. 59, 195 denotation 76
blind commitment 33 de re quantification 55, 196
blind social commitment 122-123 derived connectives 81
bold intention reconsideration 36, 40, 41 design stance 43
Bond, A. 18 desire 23, 31
branching time 55-57, 71, 72 accessibility relation 74
Bratman, M. x, 7, 45, 103 modality 50, 52
Bretier, P. 134 directive speech acts 128
Biichi automata 170 discrete time 55, 71
bullet operator 196 disjunction 48
Dogstar (grunge rock band) 54
224 Index

domain of quantification 75-76 H

doxastic logic 183
DS 1 NAS A space probe ix Haddadi, A. 161
dynamic logic 47, 82, 89 Halpern, J. Y. 12, 19, 89
Hayes, P. 89, 161
E Hendler, J. 18
Hindriks, K. 177
electronic commerce ix, 6 Hintikka, J. 89, 182, 184
Emerson, E. A. 12, 89, 171 Hoare, C. A. R. 129, 143, 165, 170
Enderton, H. 89 Huhns, M. N. 18
epistemic logic 183
equality 48
Etzioni, o. 18
Euclidean relations 74 illocutionary act 127
executable logic 15, 167-168 implication 48
existential inform speech acts 136-139
patb quantifier 57, 81 intention accessibility relation 74
quantifier 48 intentional
expressive speech acts 128 stance 42
system 42
F intention-belief
consistency 101
felicity conditions 127 incompleteness 25
Ferber, 1. 17 inconsistency 25
Ferguson, 1. A. 45 intention-desire consistency 106
filtering 31 intention 23
filter of admissibility 24 future-directed 23, 107
FIPA ACL 132-134 impossible 35
FIPA standardization effort 132 modality 50, 52
first-order logic 12-13, 181 persistence of 24
Fisher, M. 201 reconsideration 35
fixed points 117 success of 35
folk psychology 41 Internet, applications of agents on ix
Frege, G. 182, 197 intersection correspondences 97-99
fullpatbs 75 introspection axioms 192
functional systems 5 IRMA agent architecture 45
future-directed intentions 23 iteration operator 63

G J

Galliers, 1. R. 144 JACK programming language 45

game theory 11, 19 JAM! BD! system 45
GAPPS program 171 JAVA programming language 45, 125-126
Giirdenfors, P. 45 Jennings, N. 3, 18, 120, 161
Gasser, 1. 18 joint commitment 120
Geneseretb, M. 89
Georgeff, M. P. x, xi, 18, 39, 45, 89, 90, 112, K
143, 161, 175
goal-directed behaviour 4 Kaelbling, 1. 18, 171
Godel, K. 197 Kamp, H. 201
GOLOG system 177 KARO logic 67
Gough, G. 201 KD45 modal logic 74, 85, 189
Grosz, B. 161 KD modal logic 74, 86, 189
groups of agents 66, 82 KIF (agent communication language) 131
Index 225

Kinny, D. 40, 166 semantics of 116-117

Konolige, K. 45, 144, 198 mutual exclusion 61
KQML (agent communication language) 131 mutual mental states xi
Kraus, s. 161
Kripke, S. 92, 111, 184 N
KT4 modal logic 189
KT5 modal logic 189 NA SA ix
necessary truth 185
L negation 48
negotiation 6, ll, 19, 158
labeled transition system 71 nested modalities 52
Lamport, L. 202 Newell, A. 44
Lesperance, Y. 177 next time temporal operator 57, 70
Levesque, H. 66, 89, 112, 120, 124, 130, 134, Nilsson, N. 89
161, 177, 191 non-determinism 56
liar paradox 198 normal modal logic 185
Lifschitz, V. 29 Norvig, P. 10, 18
linear past time 55
liveness properties 61 o
local invariance 60
locutionary act 127 OASIS (air-traffic control system) 32
logical omniscience 191 objects 125-126
logic w regular expression 170

dynamic 47 opaque contexts 181

first-order 12-13, 181 open-minded commitment 33
in AI 11-12, 19 optimal intention reconsideration 38-39
multimodal 14, 89, 112 option generation 31
Lomuscio, A. 112 over commitment 35
,CORA (logic of rational agents) xi, 9, 14,
47-90 p
lost update problem 125
package deal problem 110
M pairwise interactions 99-100
Parsons, S. 161
Manna, Z. 170, 172 partial correctness 60
many-sorted logic 69 path 75
McCarth�J. 19, 43, 89, 161 formulae 57, 77
means-ends reasoning 21, 25 quantifiers 57
mental attitudes 41 percepts 28
mentalistic terminology 40 performative verbs 127
meta-level control 37, 40, 41 Perlis, D. 198
metalogics 197 perlocutionary act 127
Meyer, J.-J. 67 Perrault, R. 129, 134, 144
Milner, R. 143 persistence of intentions 24
minimal social convention 123-124 physical stance 43
model sets 184 Piaget, J. 43
models for ,CORA 77 plan generation 30
Montague, R. 198 plan library 30
Moore, R. 89, 144, 161 plans 8, 22, 28, 45
Morgenstern, L. 161 as recipes 28
1-£ calculus 90 Pnueli, A. 5, 169, 172, 201
Miiller, J. 45 Pnuelian reactive system 5
mutual beliefs 113-118 Pollack, M. 45
as fixed points 117-118 possible worlds 73, 92, 183
226 Index

post-conditions 29, 129 Russell, S. 10, 18

potential for cooperation 154
practical reasoning 7, 21 s
pre-conditions 29, 129
predicates 48, 69 s5 modal logic 87, 189
unary 48 Sadek, D. l33, 144
preparatory conditions 128 safety properties 60
pre-team mental state 157 satisfaction relations 77
proactiveness, agent property 3-4 for path formulae 77
pro-attitudes 23 for state formulae 78
procedural reasoning system (PRS ) 8, 45, 166 Schild, K. 90
PROLOG (programming language) 28 Schut, M. 40
proofs 15 Searle, J. 124
purposeful behaviour 4 self-reference 198
semantics
Q of action expressions 78-81
of .cORA 78-79
quantification semantic tableaux method 169
de dicta 55, 196 semantic verification 174-175
de re 55, 196 sequential composition operator 63
quantified epistemic logic 193 serendipity 36
quantifier serial relations 74
existential 48 SharedPlans formalism 161
universal 48 Shoham, Y. 42, 44
quantifYing in 54-55, 196 side effect problem 110
question protocols 141 Simon, H. 18
Singh, M. P. 18, 67, 144, 161
R single-minded commitment 33
Sistla, P. 202
Rao, A. x, xi, 18, 89, 90, 112, 161, 175 Smullyan, R. 169
rate of world change 40 social ability 3, 5
rational agents 1 software engineering 5-6, 163
reactiveness, agent property 3-4 sometime temporal operator 57, 70, 81
reactive systems 5, 170 soundness of plans 29
realism 100 specification 163
speech acts
weak: 106--110
commissive 128
reductionism 119
declarative 128
Reeves, K. 54
directive 128
referential opacity 181
expressive 128
refinement 165
theory 126--l31
Reiter, R. 144, 177
state formulae 57, 77
relational structures l3
stimulus -+ response behaviour 4
replanning 34
structural correspondence 93, 96-97
representation, symbolic 28
structural subset relation 96
representative speech acts 128
structural superset relation 96
request protocols 142
Stuart, C. 144
request speech acts l39-141
subset correspondences 93
requests that l39
substitution, logical 53
requests to l39
subworld relation 93
resource bounded agents 7, 22
synchronization 125
Rosenschein, 1. 19
syntax of .cORA 72
Rosenschein, S. 171
Rosner, R. 169
RULER program 171
Index 227

Tate, A. 18
team formation 156
teamwork 119-124
temporal implication 61
temporal logic 13, 56--61, 143, 201-203
terms, logical 70
test operator 63
Thatcher, M. 195
theoretical reasoning 21
time constrained decision making 22
total correctness 61
total relation 72
TOURINGMACHINES agent architecture 45
transitive relations 74
truth functional operators 181
truth predicate 198
Turner, R. 200

UB temporal logic 202

unary predicates 48
ungrounded semantics 164
universal
path quantifier 57
quantifier 48
universe of discourse 75
UNIX operating system 2
until temporal operator 57, 70, 81

V-z

Vardi, M. 170
variables
assignment for 76
logical 48
VDM specification language 165
verification 172
weak realism 106-110
Weiss, G. 18
Werner, E.
Wolper, P. 170, 202
Wooldridge, M. 3, 18
world change, rate of 40
z specification language 165