Proceedings of the IEEE International Symposium on Industrial Electronics, june 17-20 1996, Warsaw.

Internal Model Control Using Neural Networks.

Isabelle Rivals and Lon Personnaz
ESPCI, Laboratoire d'lectronique, 10 rue Vauquelin
75231 Paris Cedex 05, France.
E-mail: [email protected], [email protected]
IM control characteristics are the consequence of the
following properties:
(a) If the process and the controller are (input-output)
stable, and if the IM is perfect, then the control system is
stable.
(b) If the process and the controller are stable, if the IM
is perfect, if the controller is the inverse of the IM, and if
there is no disturbance, then perfect control is achieved.
(c) If the controller steady-state gain is equal to the
inverse of the IM steady-state gain, and if the control
system is stable with this controller, then offset-free
control is obtained for constant setpoints and output
disturbances.

Abstract We propose a design procedure of neural Internal

Model control systems for processes with delay. We assume
that a stable discrete-time neural model of the process is
available. We show that the design of a Model Reference
controller for Internal Model control necessitates only the
training of the inverse of the model deprived from its delay,
provided this inverse exists and is stable. As the robustness
properties intrinsic to Internal Model control systems are
only obtained if the inverse model is exact, it is also shown
how to limit the effects of a possible inaccuracy of the inverse
model due to its training. Computer simulations illustrate the
proposed design procedure.

I. INTRODUCTION
There are two ways to cope with nonlinear control
design for plants subject to uncertainty and disturbances:
robust and adapative control. While neural adapative
control is being intensively developed [1][2], the robust
technique of Internal Model control is investigated here,
using neural networks. The ability of the latter for
nonlinear black-box modeling of processes and their
inverses is exploited throughout the paper. In this
introduction, we present basic notions concerning Internal
Model control.
A control system consists of the process to be controlled
and of a control device chosen by the designer, which
computes the control input so as to convey the desired
behavior to the control system. The control device consists
of a controller and possibly other elements (observer,
filter, internal model). In this paper, the control device
imposes the desired dynamic behavior with the help of a
Model Reference controller, which is described in section
III. We distinguish between two types of control systems:
- Simple Feedback control systems: their control device is
made of the controller only. They are the most classic
ones in the field of nonadaptive neural control, but their
performances rely heavily on the accuracy of the model
used for the design of the controller.
- Internal Model (IM) control systems: they are
characterized by a control device consisting of the
controller and of a simulation model of the process, the
IM. The IM loop computes the difference between the
outputs of the process and of the IM, as shown in Fig. 1
(the control device is represented on a gray background).
This difference represents the effect of disturbances and
of a mismatch of the model. IM control devices have
been shown to have good robustness properties against
disturbances and model mismatch in the case of a linear
model of the process [3]. Developments of IM control in
the case of nonlinear models of the process have been
proposed, mainly for continous-time models [4], but also
for discrete-time models [5]; neural (discrete-time) IM
control systems are discussed in [6] [7]. Discrete-time

Effect of disturbances
r

r*

Controller

Process

yp

Internal
Model

+
y

Fig. 1. Internal Model control system.

As a consequence of c), if the controller is made of the

inverse of the IM cascaded with a low-pass filter, and if
the control system is stable, then offset-free control is
obtained for constant inputs, i.e. setpoint and output
disturbances. Moreover, the filter introduces robustness
against a possible mismatch of the IM, and, though the
gain of the control device without the filter is not infinite
as in the continuous-time case, its interest is to smooth
out rapidly changing inputs.
Our aim is to show that the design of Simple Feedback
and IM control systems with neural networks is very
similar, and that it is straightforward to improve a classic
Simple Feedback control system with poor performances,
because of disturbances or of a model mismatch, by
changing it into an IM control system.
II. NEURAL MODELS FOR CONTROL
We are interested in single input/single output processes.
We consider discrete-time deterministic nonlinear inputoutput models. The process is described by the assumed
model:
y p(k+1) = h y p{kk-n+1, u{k-d+1
(1)
k-m+1
k
where yp{k-n+1 denotes the set of past outputs of the
process yp(k), , yp(k-n+1) , and u{k-d+1
k-m+1 the set of the
past control inputs u(k-d+1), , u(k-m+1) ; d1 is the
delay of the assumed model; h is an unknown nonlinear
function. Possible disturbances are assumed not to be
measurable, hence the model does not explicitely represent
them. Nevertheless, they will appear in section V, devoted
to the simulation examples.

A neural model associated to the process assumed model

(1) is of the form:
k-d+1
y(k+1) = y{kk-n+1 , u{k-m+1
;
(2)
where is a nonlinear function implemented by a
feedforward network with weights . If the family of
functions defined by the feedforward part of the network is
rich enough with respect to the complexity of h (i.e. the
number of neurons is adequate), and if the training
algorithm is efficient, then the function implemented by
the neural network will be arbitrarily close to the function
h in the domain defined by the training sequences. The
neural identification procedure is extensively discussed in
[8], and appropriate training algorithms are described in
[9] for example.
In the case of a process with delay (d>1), we will also
need the model (2) deprived from its delay:
z(k+1) = z{kk-n+1, u{kk-m+d
(3)
is the same function than in (2), being now fixed.
III. DESIGN OF THE NEURAL CONTROLLER
In subsection A, we present the Model Reference (MR)
control objective. Then, we define the theoretical, or exact,
MR controller imposing the reference dynamic behavior,
and the strongly related theoretical inverse model. The
training of an empirical neural MR controller is then
presented in subsection D.
A. Control Objective
We consider the problem of tracking a setpoint sequence
{r(k )}, possibly in the presence of deterministic
disturbances which might occur randomly, and whose
effect must be cancelled (regulation).
The desired dynamical behavior of the control system is
chosen to be given by a stable reference model. In this
paper, the reference model is linear, but according to
known charateristics of the process and possibly to the
saturations of the controller, a nonlinear reference model
can be chosen as well (see for example the training of a
minimum-time heading controller for a 4WD Mercedes in
[10]). For process (1) with delay d and order n, a suitable
linear reference model is given by:
E (q-1) y r(k) = q-d H (q-1) r(k)
(4)
where r denotes the setpoint and y r the output of the
reference model, and where:
E (q -1) = 1 + e1q -1 + + enq -n
(5)
H (q -1) = h0 + h1q -1 + + hnq -n h0 0
-1
q is the backward shift operator, which will be omitted in
the following when it is the argument of a polynomial. The
theoretical objective is thus to impose the perfect control:
E y p(k) = q-d H r(k)
(6)

behavior of the reference model at all time k k+d, if no

disturbance occurs.
r

yr
yr
=0

Theoretical
MR Controller
(E, H, d)

Model
(delay d)

Fig. 2. Theoretical Model Reference controller.

We consider the class of models whose MR controller

exists and is stable, in the state domain of interest. The MR
controller is well known in the linear literature [11].
C. Theoretical Inverse Model
The theoretical inverse of model (2) is defined, if it
exists, as the controller which, when cascaded with model
(2), imposes the behavior of a pure delay d:
y(k+d) = r(k) = yr(k+d)
(8)
The inverse model is thus the MR controller for E = H = 1
(i.e. the reference model is the delay d). It is known in the
linear literature as the one-step-ahead controller [11].
D. Design of the Neural Reference Model Controller
As will be shown in section IV, only the case of a model
without delay (d=1) needs to be considered.
The objective is to impose the desired dynamic behavior
given by the reference model, i.e. relationship (7), from
time k+1 on. Relationship (7) can be rewritten as:
y(k+1) = H r(k) + 1 E y(k+1)
(9)
Using the model expression (2) with d=1, it follows that:
k
y{kk-n+1 , u{k-m+1
= H r(k) + 1 E y(k+1) (10)
We suppose that it is possible to express u(k) as a function
g of H r(k) + 1 E y(k+1), i.e. of the past outputs y{kk-n+1
and inputs u{k-1
k-m+1, such that (10) is verified in the domain
of interest. The theoretical MR controller can be expressed
as:
k
k-1
u(k) = g H r(k) + 1 E y(k+1), y{k-n+1
, u{k-m+1
(11)
The MR controller is of order m10. Using the same
derivation, the theoretical inverse model is given by:
u(k) = g y r(k+1), y {kk-n+1, u{k-1
(12)
k-m+1
The MR controller is thus equivalent to the inverse model
driven by the output of a feedforward system defined by
the polynomials E and H of the reference model; this
output yral(k+1) computed at time k is given by:
yral(k+1) = H r(k) + 1 E y(k+1)
(13)
This feedforward model is called a rallying model. The
equivalence is illustrated in Fig. 3:
r

B. Theoretical Model Reference Controller

Theoretical
MR Controller
(E, H, d=1)

Model
(d=1)

The theoretical MR controller of model (2) is defined, if

it exists, as the controller which, when cascaded with
model (2) in a simple feedback system, imposes the
reference dynamic behavior (4), as shown in Fig. 2:
E y(k) = q -d H r(k)
(7)
Thus, whatever the state of the model at time k, the
theoretical MR control input u(k) is such that the behavior
of the system {MR controller+model} is identical to the

Reference
Model
(E, H, d)

Theoretical MR Controller (E, H, d=1)

y ral
Rallying
Theoretical
Model
Inverse
Model
(E, H, d=1)
y

Model
(d=1)

Fig. 3. Model Reference controller equivalence for a model

without delay.

Thus, in order to obtain the MR controller, it is only

necessary to estimate the function g, i.e. to train the inverse
model.
The training system of the inverse of a model without
delay is shown in Fig. 4. It consists of:
the reference model generating the desired output
sequence; at time k, it computes:
y r(k+1) = H r(k) + 1 E y r(k+1)
(14)
the neural inverse model (to be trained); at time k, it
computes:
k
k-1
u(k) = yr(k+1), y{k-n+1
, u{k-m+1
;
(15)
where is a nonlinear function implemented by a
feedforward network with weights .
the model, whose weights are fixed; at time k, u(k) being
available, it computes:
k
y(k+1) = y{kk-n+1 , u{k-m+1
(16)
r

yr

Neural
Inverse Model
(to be trained)
Back-Prop

Reference
Model
(E, H, d=1)

yr

Model
(d=1, fixed)
Back-Prop

yr
+

Fig. 4. Training system of the inverse of a nonlinear model

without delay.

If the family of functions defined by the feedforward

part of the network is rich enough with respect to the
complexity of g, and if the training algorithm is efficient,
then the function implemented by the neural network will
be arbitrarily close to the function g in the domain defined
by the training sequences.
The MR controller is then obtained from the trained
inverse model by replacing yr(k+1) in (15) by the output of
the rallying model H r(k) + 1 E y(k+1) :
k
k-1
u(k) = H r(k) + 1 E y(k+1), y{k-n+1
, u{k-m+1
(17)
whose weights are now fixed. An advantage of the
proposed design of a MR controller is that, if the behavior
of the control system based on the MR controller is not
entirely satisfactory, it is possible to tune the rallying
model without training the MR controller anew.
Note that the design of the MR controller is more
complex for a model with delay. If d>1, a difference arises
from the fact that 1 E y (k+d) in (10) contains not only
past or present outputs, but also the future outputs y(k+1),
y(k+d-1). In that case, the MR controller must be
designed and trained as a function of the reference model
and of its delay (see [12] for a full treatment of the case
d>1).
IV. THE INTERNAL MODEL CONTROL SYSTEM
In order to show the simplicity of the design of IM
control systems, the classic Simple Feedback control
system design is first dealt with.
Simple Feedback Control System
Having trained a MR controller, the design of a Simple
Feedback control system is straightforward: the MR

controller is cascaded with the process. In the particular

case of a model without delay, the MR controller consists
of the inverse model cascaded with the rallying model
associated to the reference model, as shown in Fig. 5. The
expression of the control is obtained by replacing y by y p in
(17):
k-1
u(k) = H r(k) + 1 E yp(k+1), yp{kk-n+1, u{k-m+1
(18)
k-1
= yral(k+1), yp{kk-n+1, u{k-m+1
where yral(k+1) is the output of the rallying model.
Neural MR Controller (d=1)
y ral
Rallying
Neural
Inverse
Model
Model
(E, H, d=1)
yp

r
yp

yp

Process

Fig. 5. Simple Feedback control system of a process without

delay.

As already mentionned in section I, such a control system

is intrinsically not robust towards a model mismatch and
disturbances, which lead to important steady-state errors.
Internal Model Control System
The controller must now control the IM. Let us show
that the case of a model with delay is dealt with very
simply. Model (2) is used as IM; let it be split into:
the delay-deprived model (3):
k
z(k+1) = z{kk-n+1, u{k-m+d
a delay (d1): y(k+d) = z(k+1).
What is needed is the MR controller for the delay-deprived
model: since the latter is a model without delay, it can be
built from:
the inverse of the delay-deprived model;
a rallying model for the delay-deprived model, which is
thus defined by the polynomials E and H of the reference
model, but with delay d=1. Its input are the IM setpoint,
and the past outputs of the delay-deprived model.
The overall IM control system is shown in Fig. 6.
r
+

eIM

MR Controller of the Delay-Deprived Model

zral
Rallying
Neural Inverse
Model
of the Delay(E, H, d'=1)
Deprived Model
z
z

r*

yp

Process
Internal Model

DelayDeprived
Model

yp
+

q-(d-1)

eIM

Fig. 6. Internal Model control system of a process with delay.

When u(k-1) becomes available, z(k) is computed by the

delay-deprived model :
k-1
z(k) = z{k-1
k-n, u{k-m+d-1
Then, let us consider the computations performed at time
k, which is defined as the instant when yp(k) is available:
a) the difference between the outputs of the process and of
the IM: eIM(k) = y p(k) y (k)
b) the corrected setpoint for the IM: r*(k) = r(k) eIM(k)
c) the output of the rallying model:
zral(k+1) = H r*(k) + 1 E z(k+1)
(19)
d) the control value, output of the inverse of the delaydeprived model:
k
k-1
u(k) = zral(k+1), z{k-n+1
, u{k-m+d
(20)
Properties of the proposed IM control system
The proposed control system has the basic properties
described in section I:

If the inverse is exact, the output of the delay-deprived

model is equal to the output of the rallying model:
z(k+1) = H r*(k) + 1 E z(k+1)
If the IM is perfect, the dynamic behavior of the control
system is thus identical to the reference behavior in the
absence of disturbances.
If only the inverse model is exact, and if the overall
control system is stable, the effect of constant
disturbances and of a model mismatch are cancelled in
the steady-state, which is an important advantage over a
Simple Feedback control system.
The rallying model of the MR controller plays the role of
the IM filter in the classical IM control system [4] which
ensures the robustness of the stability of the control
system with respect to a process mismatch, and the
smoothness of the control. With our method, the design
of the filter is straightforward.
Design precautions
Let us now specify the design precautions that must be
taken in the case of a neural network based IM control
system in order to maximize its robustness properties. One
must be aware that the validity domain of the neural
inverse model is restricted to the region of the state space it
was trained in. This training region must be larger than the
operating region desired for the process since, during the
operating phase, the neural inverse model controls the IM,
which might evolve in a larger region of the state space
than the process, due to a model mismatch and to
disturbances.
Remarks
Neural IM control system are often simplified in the two
following ways:
Instead of the feedforward rallying model, the feedback
reference model with input r* and output zr* is used to
drive the inverse model, as in [7] for example. This
assumes that z(k) = zral(k) k holds, i.e. that the inverse
model is exact. But, dealing with general nonlinear
neural models, this is usually not true. As a matter of
fact, we showed in [13] that an IM control system with a
rallying model is more robust towards a controller/model
mismatch than with the reference model.
The IM is removed, in [6] for example, as in Fig. 7:
r

r*
+

eIM

Reference
Model
(E, H, d'=1)

zr *

Neural Inverse
of the DelayDeprived Model

Process

q -d

systems are designed for these processes, and compared to

Simple Feedback control systems on representative
setpoint sequences, in the presence of disturbances.
A. Process 1 (without delay)
Process 1 is simulated by:
yp(k+1) = h yp(k), yp(k-1), u(k)
2
(21)
24+yp(k)
=
yp(k) 0.8 u(k) 2 yp(k-1) + 0.5 u(k)
30
1+u(k)
Analysis of the Process Behavior
For large input amplitudes, the process is a stable
nonlinear oscillatory low-pass filter (see Fig. 8). More
specifically, around operating points {y 0 , u0 }, its behavior
is the following: (i) around {0, 0}, it behaves like a
damped first-order filter with time constant 3.6 T (T is the
sampling period); (ii) around operating points
corresponding to 0.1 < |u 0 | < 0.5, it behaves like an
oscillatory second-order filter with pseudo-period 5 T; (iii)
for operating points corresponding to larger u0 , the
behavior becomes more complex. The amplitude of u is
limited to [-5; 5] by a saturation of the actuator output; for
a constant amplitude 5 of the control input, the output is
equal to 2.7, and for an amplitude of -5, to -2.3.
Reference Model
The reference regulation and tracking dynamic behavior
is chosen to be given by a damped linear second-order
filter with time constants T and 1.2 T, and with unit static
gain:
yr(k+1) + e1 yr(k) + e2 yr(k-1) = h0 r(k) + h1 r(k-1) (22)
where: e1 = 0.803; e2 = 0.160; h0 = 0.232; h1 = 0.126.
1) Modeling of the Process
The training control input sequence consists of pulses of
random amplitude in the range [-5; 5] with a duration of 10
sampling periods; the total length of the sequence is 1000.
The control input sequence used for the performance
estimation is generated in the same manner, and is shown
in Fig. 8.
u

yp
yp
+
yr*

-5

yp
0

200

400

600

800

1000

Fig. 8. Performance estimation sequences for modeling.

eIM

Fig. 7. Simplified Internal Model control system to avoid.

As a matter of fact, the process is usually equipped with

an actuator whose output is limited in amplitude. The
existence of an inverse is then restricted to these limits,
and so the legimity of the IM removal. As a
consequence, the control system of Fig. 7 is not immune
to wind-up (the loop with the delay d creates an
integrator that is never switched off).
V. ILLUSTRATIVE EXAMPLES

We retained the following candidate:

y(k+1) = y(k), y(k-1), u(k)
(23)
where is a function implemented by a fully connected
network1 with 5 sigmoidal hidden neurons and a linear
output neuron. Using a quasi-Newton algorithm, a
training mean square error (TMSE) of 1.2 10-5 and a
1 A fully connected network with Ne inputs x 1 , , xN e , and Nn
ordered neurons, is such that the output xi of neuron i is given by:
i-1

x i = fi

ij x j

i = Ne+1 to Ne+Nn

j=1

This section presents the modeling and control of two

simulated processes, without and with delay. IM control

where f i is the activation function of neuron i. Neuron Ne+Nn is

the output neuron.

performance mean square error, i.e. a MSE on the

sequences for performance estimation (PMSE), of 1.4 10-5
were obtained.
2) Training of the Inverse Model
We assume that the process is to be controlled for
setpoint pulses of amplitude in the range [-2; 2.5]. The
inverse of model (23) is trained using the training system
described in section III.D Fig. 4; it consists of:
the reference model (22) generating the desired output
sequence yr(k) from the setpoint sequence r(k) .
the neural inverse model (to be trained):
u(k) = yr(k+1), y(k), y(k-1);
(24)
where is a function implemented by a fully connected
network with sigmoidal hidden neurons and where the
activation function of the output neuron is the saturation
function between [-5; 5] (which corresponds to the
actuator limits).
the neural model (23), whose weights are now fixed.
The training setpoint sequence {r(k)}consists of pulses of
random amplitude in the range [-2.5; 3] with a duration of
20 sampling periods; the total length of the sequence is
1000. The maximum amplitudes are chosen so as to fully
explore the desired output amplitude range [-2; 2.5] and to
encounter the saturations. With 5 hidden neurons, a TMSE
of 5.6 10 -4 is obtained. Since most errors are due to the
control saturation, adding neurons does not improve the
TMSE significantly.
3) Control of the Process
The process is affected by a randomly occuring pulseshaped disturbance of amplitude 0.5, denoted by da, acting
additively on the output:
x p(k+1) = h x p(k), x p(k-1), u(k)
(25)
y p(k) = x p(k) + da(k)
The setpoint sequence consists of pulses with amplitudes
in the range [-2; 2.5]. The Simple Feedback control system
is first tested.
Simple Feedback Control System
The behavior of the Simple Feedback control system is
shown in Fig. 9:

disturbance occurs, there is a small steady-state error, due

to a controller inaccuracy. As a matter of fact, the value of
the output of the model reaches 3, that is the limit of the
domain the controller was trained in. Retraining the
controller in a larger domain leads to zero error, but our
goal was to emphasize the importance of the design
precautions sketched in section IV.
5

-5
0

50

100

150

200

250

300

Fig. 10. Internal Model control of process 1 (thin line: r; dotted

line: da ; dashed line: u; thick line: yp ; thick dotted line: y).

B. Process 10 (with delay)

In order to show the similarity of the design in the case
of a process with delay, the second process is chosen to be
characterized by the same input-output relationship than
process 1, but with a delay d=10. Process 10 is thus
simulated by the following discrete-time system:
x p(k+1) = h x p(k), x p(k-1), u(k-9)
(26)
y p(k) = x p(k) + da(k)
We choose to deduce a model of (26) from the model of
process 1 (23) by adding a delay of 9 time steps to it:
y(k+1) = y(k), y(k-1), u(k-9)
(27)
As shown in section IV for models with delay, the neural
controller (24) can be used in the IM control system for
process 10 (but not in a Simple Feedback control system,
we will therefore not make the comparison for this
process). Its behavior is shown in Fig. 11:
5

5
-5
0

50

100

150

200

250

300

Fig. 11. Internal Model control of process 10 (thin line: r;

dotted line: da; dashed line: u; thick line: yp ; thick dotted line: z).

-5
0

50

100

150

200

250

300

Fig. 9. Simple Feedback control of process 1 (thin line: r ;

dotted line: da; dashed line: u; thick line: yp ).

A disadvantage of Simple Feedback control appears

clearly: constant disturbances lead to important steadystate errors.
Internal Model Control System
The behavior of the IM control system is shown in Fig.
10. y denotes the output of the IM. Thanks to the IM
control structure, the performance in the steady-state is
greatly improved (it appears clearly in the control error
shown in Fig. 12). Nevertheless, the second time the

z denotes the output of model (27) deprived from its delay.

If we except the uncancellable errors due to the pure delay,
the performance of the control system of process 10 is
almost identical to that of process 1.
1

-1

50

100

150

200

250

300

Fig. 12. Comparison of the control errors (thin dashed line: da;
thin continuous line: Simple Feedback control of process 1;
thick dotted line: Internal Model control of process 1; thick line:
Internal Model control of process 10).

The errors of the three control systems are shown in Fig.

12. The control error is computed as the difference
between the output of the reference model (not shown on
the graphs for the sake of clarity) and of the process.
VI. CONCLUSION
We have presented a design procedure of Internal Model
control systems for the problem of tracking a setpoint
sequence with the dynamic behavior given by a reference
model. We have shown that the design of the Model
Reference controller can be performed independantly from
the delay of the process, since it can be built of the inverse
of the model deprived from its delay (trained), and of a
rallying model associated to the reference model.
Moreover, the control system using the Model Reference
controller can be modified by tuning the rallying model
(not trained), if necessary. The advantage of Internal
Model control systems is their robustness with respect to a
model mismatch and to disturbances. Nevertheless, these
properties are only obtained if the Model Reference
controller is close to the theoretical one; we have thus
stressed the design precautions that must be taken in order
to avoid the consequences of a possible controller/model
mismatch. The proposed Internal Model design procedure
was successfully applied to the piloting (velocity control)
of a 4WD Mercedes vehicle [14]. We are now
investigating the use of affine black-box models, which
allow a direct and thus exact design of their inverse model
[12].
VII. REFERENCES
[1] Sanner R., Slotine J.-J., Stable adaptive control of
robot manipulators using neural networks, Neural
Computation Vol. 7 N4, 1995, pp. 753-790.
[2] Narendra K. S., Mukhopadhyay S., Adaptive control
of nonlinear multivariable systems using neural
networks, Neural Networks Vol. 7 N5, 1994, pp.
737-752.
[3] Morari M., Zafiriou E., Robust process control,
Prentice-Hall International Editions, 1989.
[4] Economou G. C., Morari M., Palsson B. O., Internal
Model control. 5. Extension to nonlinear systems,
Ind. Eng. Chem. Process Des. Dev. Vol. 25, 1986, pp.
403-411.
[5] Alvarez J., An internal model controller for
nonlinear systems, Proceedings of the 3rd European
Control Conference, Rome, Italy, September 1995,
pp. 301-306.
[6] Psichogios D. C., Ungar L. H., Direct and indirect
model based control using artificial networks, Ind.
Eng. Chem. Res. Vol. 30 N12, 1991, pp. 2564-2573.
[7] Hunt K., Sbarbaro D. Neural networks for nonlinear
internal model control, IEE Proc.-D Vol. 138 N5,
1991, pp.431-438.
[8] Rivals I., Personnaz L., Black-box modeling with
state-space neural networks, to appear in Neural
Adaptive Control Technology I, R. Zbikowski and K.
J. Hunt eds., World Scientific, 1996.
[9] Nerrand O., Roussel-Ragot P., Personnaz L., Dreyfus
G., Neural networks and nonlinear adaptive filtering:

unifying concepts and new algorithms, Neural

Computation Vol. 5 N2, 1993, pp. 165-199.
[10] Rivals I., Personnaz L., Dreyfus G., Canas D., Realtime control of an autonomous vehicle: a neural
network approach to the path following problem, 5th
International Conference on Neural Networks and
their Applications, Nmes, France, 1993, pp. 219-229.
[11] Goodwin G. C., Sin K. S., Adaptive filtering
prediction and control, Prentice-Hall, New Jersey,
1984.
[12] Rivals I., Personnaz L., Model Reference control
using neural networks, 1996, submitted for
publication.
[13] Rivals I., Modlisation et commande de processus
par rseaux de neurones ; application au pilotage
dun vhicule autonome, Thse de Doctorat de
lUniversit Paris 6, 1995.
[14] Rivals I., Canas D., Personnaz L., Dreyfus G.,
Modeling and control of mobile robots and
intelligent vehicles by neural networks, IEEE
Conference on Intelligent Vehicles, Paris, 24-26
octobre 1994, pp. 137-142.