1392 IEEE TRANSACTIONS ON SYSTEMS. MAN. AND CYBERNETICS, VOL. 23, NO.

5, SEPTEMBER/OCTOBER 1993

Fuzzy Gain Scheduling of PID Controllers

Zhen-Yu Zhao, Member, IEEE, Masayoshi Tomizuka, Member, IEEE, and Satoru Isaka, Member, IEEE

Abstract- This paper describes the development of a fuzzy trol actions. Although they do not have an apparent struc-
gain scheduling scheme of PID controllers for process control. ture of PID controllers, fuzzy logic controllers may be
Fuzzy rules and reasoning are utilized on-line to determine the
considered nonlinear PID controllers whose parameters
controller parameters based on the error signal and its first
difference. Simulation results demonstrate that better control can be determined on-line based on the error signal and
performance can be achieved in comparison with Ziegler- their time derivative or difference.
Nichols controllers and Kitamori's PID controllers. In this paper, a rule-based scheme for gain scheduling
of PID controllers is proposed for process control. The
new scheme utilizes fuzzy rules and reasoning to deter-
I. INTRODUCTION
mine the controller parameters, and the PID controller
HE BEST-KNOWN controllers used in industrial
T control processes are proportional-integral-derivative
(PID) controllers because of their simple structure and ro-
generates the control signal. It is demonstrated in this pa-
per that human expertise on PID gain scheduling can be
represented in fuzzy rules. Furthermore, better control
bust performance in a wide range of operating conditions. performance can be expected in the proposed method than
The design of such a controller requires specification of that of the PID controllers with fixed parameters.
three parameters: proportional gain, integral time con-
stant, and derivative time constant. So far, great effort has II. PID CONTROLLER
been devoted to develop methods to reduce the time spent
on optimizing the choice of controller parameters [8], The transfer function of a PID controller has the fol-
[15]. The PID controllers in the literature can be divided lowing form:
into two main categories. In the first category, the con- Gc(s) = «, + Ki/s + Kds (1)
troller parameters are fixed during control after they have where Kp , Ki, and Kd are the proportional, integral, and
been tuned or chosen in a certain optimal way. The derivative gains, respectively. Another useful equivalent
Ziegler-Nichols tuning formula is perhaps the most well- form of the PID controller is
known tuning method [5], [19]. Some other methods exist
for the PID tuning (see e.g., [1], [6], [7]). The PID con- + l/(Tis) + Tds)
Gc(s) = Kp (1 (2)
trollers of this category are simple, but cannot always ef- where T, = Kp/ K, and Td = Kd/ K p. T, and Td are known
fectively control systems with changing parameters, and as the integral and derivative time constants, respectively.
may need frequent on-line retuning. The controllers of the The discrete-time equivalent expression for PID control
second category have a structure similar to PID control- used in this paper is given as
lers, but their parameters are adapted on-line based on
n K
parameter estimation, which requires certain knowledge
of the process, e.g., the structure of the plant model [2],
u(k) = Kpe(k) + «, T, i~1 e(i) + T: l1e(k).
[17]. Such controllers are called adaptive PID controllers
Here, u (k) is the control signal, e (k) is the error between
in order to differentiate them from those of the first cate-
the reference and the process output, T, is the sampling
gory. period for the controller, and l1e(k) ~ e(k) - e(k - 1).
The application of knowledge-based systems in process
control is growing, especially in the field of fuzzy control
The parameters of the PID controller Kp , s; and or «,
K p , Ti, and Td can be manipulated to produce various re-
[9], [10], [12]-[14]. In fuzzy control, linguistic descrip-
sponse curves from a given process. Finding optimum ad-
tions of human expertise in controlling a process are rep-
justments of a controller for a given process is not trivial.
resented as fuzzy rules or relations. This knowledge base
In the following section, an on-line gain scheduling
is used by an inference mechanism, in conjunction with
scheme of the PID controller based on fuzzy rules is in-
some knowledge of the states of the process (say, of
troduced.
measured response variables) in order to determine con-

Manuscript received April 10, 1992; revised October 3, 1992. III. Fuzzy GAIN SCHEDULING
Z.-Y. Zhao was with the Department of Mechanical Engineering, Uni-
versity of California at Berkeley, Berkeley, CA 94720, and is currently Fig. 1 shows the PID control system with a fuzzy gain
with Omron Advanced Systems, Inc., in Santa Clara, CA. scheduler. The approach taken here is to exploit fuzzy
M. Tomizuka is with the Department of Mechanical Engineering, Uni- rules and reasoning to generate controller parameters.
versity of California at Berkeley, Berkeley, CA 94720.
S. Isaka is with Omron Advanced Systems, Inc., Santa Clara, CA 95054.
It is assumed that Kp , K; are in prescribed ranges
IEEE Log Number 9209681. [Kp.min' Kp,max] and [Kd . min' K p,max], respectively. The ap-
0018-9472/93$03,00 © 1993 IEEE
ZHAO et al.: FUZZY GAIN SCHEDULING OF PID CONTROLLERS 1393

NB NM NS ZO PS PM PB

Input

+
x=e(k) or t.e(k)
Fig. 2. Membership functions for e(k) and lie (k).
Fig. 1. PID control system with a fuzzy gain scheduler.

propriate ranges are determined experimentally and will

be given in equation (12). For convenience, Kp and Kd are
normalized into the range between zero and one by the
following linear transformation:

K; = (Kp - Kp.min) / (Kp. max - Kp.min)

K d = (Kd - Kd.min)/(Kd.max - Kd . min ) . (3)

In the proposed scheme, PID parameters are determined

based on the current error e (k) and its first difference
t:.. e (k). The integral time constant is determined with ref- x=K~ or K
erence to the derivative time constant, i.e.,
d
Fig. 3. Membership functions for K~ and K~.
T; = ex.Td (4)
and the integral gain is thus obtained by Output

K; = Kp/(ex.Td) = K~/(ex.Kd)' (5)

The parameters K;, K d, and ex. are determined by a set of
Setpoint ~-Of----'".-'--.",.....""",,=---------
fuzzy rules of the form

if e(k) is A; and t:..e(k) is B;, then K; is C;, K d is D;,

and ex. = a,
i = 1,2, ... ,m. (6)
Here, A;, B;, C;, and D; are fuzzy sets on the correspond-
a11l.----~---~---~---~----
ing supporting sets; a, is a constant. The membership
TIME
functions (MF) of these fuzzy sets for e (k) and t:.. e (k) are
shown in Fig. 2. In this figure, N represents negative, P Fig. 4. Process step response.
positive, ZO approximately zero, S small, M medium, B
big. Thus NM stands for negative-medium, PB for posi-
tive big, and so on. ample of a desired time response. At the beginning, i.e.,
The fuzzy sets C; and D; may be either Big or Small around at, we need a big control signal in order to achieve
and are characterized by the membership functions shown a fast rise time. To produce a big control signal, the PID
in Fig. 3, where the grade of the membership functions fl controller should have a large proportional gain, a large
and the variable x (= K; or K d) have the following rela- integral gain, and a small derivative gain. Thus the pro-
tion: portional gain (K;) can be represented by a fuzzy set Big,
and the derivative gain K d by a fuzzy set Small. The in-
flSMALL (x) - ~ In x or XSMALL (fl) = e -41' tegral action is determined with reference to the derivative
action as in (4). For the PID controller, taking a small ex.
for Small, or a small integral time constant T; will result in a strong
- ~ In (1 - x) or integral action. Whether the integral action should be
flBIG (x) XBIoCfl)
strong or weak is determined in the scheme by compari-
for Big. (7) son with the well-known Ziegler-Nichols PID tuning rule.
In the Ziegler-Nichols rule, the integral time constant T;
The fuzzy rules in (6) may be extracted from operator's is always taken four times as large as the derivative time
expertise. Here we drive the rules experimentally based constant. That is, ex. is equal to 4. In the proposed scheme,
on the step response of the process. Fig. 4 shows an ex- ex. takes a value less than 4 (say 2) to generate a "stronger"
1394 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 23, NO. S, SEPTEMBER/OCTOBER 1993

integral action, Therefore, the rule around Q\ reads

if e(k) is PB and t1e(k) is ZO, then K~ is Big, K d is

Small, and ex = 2, S MS M B

Note that ex may also be considered as a fuzzy number

which has a singleton membership function as shown in
Fig, 5, For example, ex becomes 2 when ex is Small.
Around point b, in Fig, 4, we expect a small control
signal to avoid a large overshoot. That is, the PID con-
troller should have a small proportional gain, a large de-
-
rivative gain, and a small integral gain, Thus the follow- 2 3 4 5 a
ing fuzzy rule is taken:
S: Small, MS: Medium Small
if e(k) is ZO and t1e(k) is NB, then K~ is Small, K d is M: Medium, B:Big
Big, and ex = 5,
Fig. 5. Singleton membership functions for 0'.

Thus a set of rules, as shown in Table I, may be used to

adapt the proportional gain (K ~). The tuning rules for TABLE I
K d and ex are given in Tables II and III, respectively. In Fuzzy TUNING RULES FOR K;,
the tables, B stands for Big, and S for Small.
tle (k)
The truth value of the ith rule in (6) lJ.i is obtained by
the product of the MF values in the antecedent part of the NB NM NS ZO PS PM PB
rule:
NB B B B B B B B
NM S B B B B B S
(8) NS S S B B B S S
e(k) ZO S S S B S S S
PS S S B B B S S
where IJ.Ai is the MF value of the fuzzy set Ai given a value PM S B B B B B S
of e (k), and IJ.Bi the MF value of the fuzzy set B, given a PB B B B B B B B

value of t1e(k).
Based on lJ.i' the values of K~ and K d for each rule are
determined from their corresponding membership func- that are due to (3) and (5):
tions. The implication process of a fuzzy rule is shown in (lla)
Fig. 6.
By using the membership functions in Fig. 2, we have (lIb)
the following condition [18]:
(lIe)

2:
i= I
lJ.i = I. (9) Based on an extensive simulation study on various pro-
cesses, a rule of thumb for determining the range of Kp
Then, the defuzzification yields the following: and the range of Kd is given as
Kp.min = O.32K u ' Kp,max = O.6Ku
(lOa) (12)

where K; and T; are, respectively, the gain and the period

of oscillation at the stability limit under P-control [19].
s; = 2: lJ.iKd.i
i= I
(lOb)
Note that there are other forms for the fuzzy tuning rules
in (6). Some examples are as follows:
m
ex = 2: lJ.iexi' (lOc) I) if e(k) is Ai and t1e(k) is B i, then K~ is Ci , K d is D i ,
i= I
and K; is E,
Here K~.i is the value of K~ corresponding to the grade u, 2) if e (k) is Ai and t1e (k) is Bi, then K; is Ci, T d is Di,
for the ith rule (see Fig. 6). K d.i is obtained in the same and t; is E,
way.
Once K~, K d, and ex are obtained, the PID controller 3) if e (k) is Ai and t1 e (k) is Bi, then u (k) =
parameters are calculated from the following equations K~ e(k) + (K;o Ts) Ej e(j) + (K~o/Ts) t1e(k)
ZHAO et al.: FUZZY GAIN SCHEDULING OF PID CONTROLLERS 1395

TABLE II
Fuzzy TUNING RULES FOR K J

tle(k)

NB NM NS ZO PS PM PB

NB S S S S S S S
NM B B S S S B B
NS B B B S B B B
e(k) ZO B B B B B B B
PS B B B S B B B
PM B B S S S B B
PB S S S S S S S

TABLE III cal entity like a supervisor is desired to monitor the per-
Fuzzy TUNING RULES FOR a
formance of the control system. Instability is detected
tle(k) preferably in an early stage if the system is unstable. Once
stability is identified during process monitoring, certain
NB NM NS ZO PS PM PB
corrective action is taken. For example, the controller pa-
NB 2 2 2 2 2 2 2 rameters are switched to a set of known stabilizing param-
NM 3 3 2 2 2 3 3 eters that guarantees that the control system will remain
NS 4 3 3 2 3 3 4
stable; or the system is shut down by setting Kp to zero if
e(k) ZO 5 4 3 3 3 4 5
PS 4 3 3 2 3 3 4 necessary .
PM 3 3 2 2 2 3 3 There are several practical methods available to iden-
PB 2 2 2 2 2 2 2
tify instability. Anderson et al. [1] suggest monitoring the
magnitude of peaks and the system is determined to be
unstable when peaks are growing in magnitude for three
Ai peaks in row. Nesler [11] uses the ratio of short-term av-
erage of the error to that of the absolute value of the error
~Ai to detect instability. Gertler and Chang [3] propose an
instability indicator by observing the output. It is also
possible to combine the above quantitative indexes with
e(k) e rule-based logic to make more accurate and reliable de-
cisions.
In some situations, a hybrid controller may be more
J.l Bj K' .
P,I K'P change and
useful reliable. For example, if a large set point
is made, the fuzzy gain scheduling scheme is first
employed to yield a fast transient response. When the er-
ror e (k) is small, i.e., the output is settling to the set point,
the scheme is then switched to a fixed gain PID controller.
M(k) ~e The stability of the closed-loop control system can be
Fig. 6. Implication process of a fuzzy rule. guaranteed while maintaining some level of control.
Sometimes a set of PID parameters are good for set point
responses but may be inappropriate for load disturbance
where K'po, K:o, K do are constant. Although these rules rejection [4], [16]. Hang [4] finds that the heuristic
have different forms, they are equivalent to each other un- Ziegler-Nichols design rule gives a set of PID parameters
der certain conditions. It seems relatively easier to set the that are good at load disturbance rejection. Therefore, a
fuzzy tuning rules by (6). PID controller with a set of fixed parameters that are ob-
tained by the Ziegler-Nichols rule can be used after the
transient stage of the process response.
IV. STABILITY

Since the parameters of the present PID controller are V. SIMULATION RESULTS AND DISCUSSIONS
functions of time, it is very difficult to analyze the stabil-
ity of the closed-loop control system. Even if the asymp- The fuzzy gain scheduling scheme has been tested on a
totic stability is assured, wild start-up transients may be variety of processes. Table IV shows the representative
intolerable in many applications. Therefore, a hierarchi- simulation results of the following second-, third-, and
1396 IEEE TRANSACTIONS ON SYSTEMS. MAN. AND CYBERNETICS. VOL. 23, NO.5, SEPTEMBER/OCTOBER 1993

Output y Output Y
1,4r---~-~:------~----------, 1.2..---......,..---------,----,-----,-----,
-r-, I'
,',\
,
1.2 ........ ·~ .: .... ...:-\- ...\./:..,.,~<:~.;>. ....-:~......----.----;-----+---___l
I: ,
" , I •

I ....... :: ';:':::'~~"""""::-:-_~"l""-~--~----:---~----l 0.8

I

0.8 ' ; .......... .; . .

.
,: .; , .
.
.;
. ;
..
r. : 0.6·
0.6 ~ ~ , . ., .. ,., .. ,:~.,. ~PI.OpOSe~,.,
0.4 . .
.: .-::.. :-:-:,. ,:~~!g.l~~~~,i~~ql.~ ...
~

0,4 .......... ;.......... ,... Proposed

0.2 .. ..Z.~i~l~r~~.i~h.9.1~ .. ,
0.2 ; , .. , ,:.~:.:.:, :.~!~~,rn~.r.i~ , .
I :
°O~----::---~-_....:.._-..:....-_..:...- __.:-..._.....J 00" -- - --:..- - - -'--- - --'- - - -'-- - ---,' -- - ......
2 4 6 10 12 14 10 15 20 25 30
TIME (sec) TIME (sec)
Fig. 7. Comparison of step responses of the controlled second-order pro- Fig. 8. Comparison of step responses of the controlled third-order pro-
cess. cess.

fourth-order processes:
Output y
e -0.55 1.4r-----:-----,-----,------:-------,
G) (s) = -(s-+-I-:;:)2 (13a) , .. ",,:,
1.2 / :.'\ , ;..
I . \

4.228
(13b) ....,/ .... -.':.>\.~..,.:~.~:'i~.,-;~~ .-=:;.~-~~--..-=" ......~--'---~-1
(s + 0.5)(S2 + 1.64s + 8.456)
0.8

G s = 27 (13c) 0.6 .. ............................ .. .. p'r()PP.s~d.."

3() (s + l)(s + 3)3'
~

0.4 . zeigl~r~~i~hOIS
In the table, Yo; represents the percent maximum over- :. • • • Kitamon
shoot, T5 stands for the 5 percent settling time, and IAE, 0.2 '
ISE are the integral of the absolute error and the integral
of the squared error, respectively [8]. The time responses 2 4 6 10
are plotted in Fig. 7, Fig. 8, and Fig. 9, respectively. The TIME (sec)
results obtained by using Ziegler-Nichols PID controllers Fig. 9. Comparison of step responses of the controlled fourth-order pro-
and Kitamori's PID controllers are also presented for cess.

comparison. The parameters of Ziegler-Nichols PID con-

trollers are determined as K p = 0.6Ku , T, = 0.5Tu , and
Ziegler-Nichols controller could be chosen slightly larger
Td = 0.125Tu ; while those of Kitamori's controllers are
in order to obtain a smaller maximum overshoot. It is also
obtained by partial model matching method [7]. A brief
found that the proposed controller is as good as or better
description of the method is given in the Appendix. Fig.
than the delicately tuned PID controller of Kitamori's,
10 shows the PID parameters determined by the fuzzy gain
which needs much more information on the plant.
scheduling scheme for controlling the second-order pro-
cess (13a). The determination process is as follows. First
the error and its first difference are calculated from the VI. CONCLUSIONS
sampled process output. Then the values of K ~ and K d for The proposed gain scheduling scheme uses fuzzy rules
each rule are determined by the fuzzy reasoning process, and reasoning to determine the PID controller parameters.
as shown in Fig. 6. Finally, the PID parameters are ob- Human knowledge and experience in control system de-
tained by using (10) and (11). sign is exploited in the tuning of a PID controller. The
The above simulations show that a variety of processes scheme has been tested on various processes in simula-
can be satisfactorily controlled by the fuzzy gain sched- tion, and satisfactory results are obtained. A supervisory
uling PID controller. It yields better control performance level may also be included to monitor the stability of the
than the Ziegler-Nichols controller does, which is con- controlled system, and a hybrid controller, i.e., rule
firmed by comparing performance indexes such as the based, plus fixed gain is possible. Although a rule of
percent maximum overshoot, the settling time, IAE, and thumb for choosing the ranges for Kp and Kd was obtained
ISE. It seems that the derivative time constant Td for the experimentally in (12), it is still possible to make further
ZHAO et al: FUZZY GAIN SCHEDULING OF PID CONTROLLERS 1397

TABLE IV
SUMMARY OF SIMULATION RESULTS

Process Zeigler-Nichols PID Proposed PID Controller

Controller Kitamoris PID Controller"

K; = 2.808 Yo, = 32% K I , = 2.212 Yo, = 6.8% Yo, = 6.0%

T, = 1.64 T, = 4.16 T, = 2.039 T, = 2.37 T, = 3.09
Td=0.41 IAE = 1.37 T, = 0.519 IAE = 1.04 IAE = !.I8
ISE = 0.871 ISE = 0.805 ISE = 0.772
K p = 2.19 Yo, = 17% Yo, = 6.1 %
T, = 1.03 T, = 5.45 T, = 5.01
T; = 0.258 IAE = 0.99 IAE = 1.01
ISE = 0.526 ISE = 0.533
G 3 (s) K; = 3.072 Yo, = 32.8% KI' = 2.357 Y,,, = 10.9% Yo, = 1.9%
T, = 1.352 T, = 3.722 T,=I.649 T, = 2.3 T, = 2.632
Td = 0.338 IAE=1.13 TJ = 0.414 IAE = 0.833 IAE = 0.811
ISE = 0.628 ISE = 0.596 ISE = 0.537

"The PID parameters of the Kitamori's controller are not available for the process G, (s) because it has
a small damping ratio, 0.282. Yo, is the percent maximum overshoot, T, is the 5 percent settling time. and
IAE. ISE are the integral of the absolute error and the integral of the squared error. respectively.

PID Parameters Choose a reference model as

(ao = al = 1)
where a is a time-scaling factor, which is chosen as small
as possible. A set of a's standard values is given
{a;} = {l, 1,0.5,0,15,0.03, , .. }
that would result in a time response with adequate damp-
ing characteristics and a short rise-time, The closed-loop
transfer function of the PID controlled system is then
4 6 10 12 14 given
TIME (sec)
Fig. 10. PID parameters of the fuzzy gain scheduler for the control of the W(s)
second-order process. 1 + I/(P(s)C(s)) 1 + sia' Ie)
where
performance improvements by fine tuning the ranges as a' ab+a;+a2s2+
well as by modifying the tuning rules in Table I through
Table III. These points require further research and de-
velopment. Equating GM(s) to W(s) gives
e = sa' I (a - 1),
ApPENDIX Expanding the right side of the above equation in s yields
KITAMORI'S PID CONTROLLER

Assume that the plant is described by the following

e
transfer function:

pes)
ab + a;s + a2s2 + , ..
))
+ a-Cal - (3) J' S-

3 a~ 2 a;
and the PID controller is + [a aa2 ----7 + a (a~ + a 3 ( 2a2a3
(3) ---;
ab ao ao
,
Co + CIS + C2S-
C(s)
S
1398 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS. VOL. 23, NO. S, SEPTEMBER/OCTOBER 1993

where ao = al = 1, By matching the coefficient in s from [15] Y. Takahashi, M. J. Rabins, and D. M. Auslander, Control and Dv-
low order to high order, we obtain namic Systems. Menlo Park, NJ: Addison-Wesley, 1970. .
[16] J. G. Truxal, Automatic Feedback Control System Synthesis. New
York: McGraw-Hill, 1955.
Co ao/ a [17] T. Yamamoto, S. Omatu, and H. Ishihara, "A construction of self-
tuning PID control system," Trans. SICE Japan., vol. 25, pp. 39-
aO(a;/ao 45.1989 (in Japanese).
118] Z. Y. Zhao, M. Tomizuka, and S. Sagara, "A fuzzy tuner for fuzzy
C2 aO{a2lao aa2a;/aO + a2(a~ - (3)}/a logic controllers," in Proc. 1992 Amer. Control Conf'., Chicago, IL,
June 24-26, 1992, pp. 2268-2272.
where a is obtained by choosing the smallest positive root [19] J. G. Ziegler and N. B. Nichols, "Optimum settings for automatic
of the following equation: controllers," Trans. ASME, vol. 64, pp. 759-768, 1942.

Zhen-Yu Zhao (M'91) was born in Jiangsu pro-

+ a3(2a2a3 - a~ - (4) = O. vince, China, on April 27, 1963. He studied at the
Department of Automation, Tsinghua University,
Thus, the PID parameters are obtained as follows: Beijing, and received the B. Eng. degree in con-
trol engineering in 1984. He received the M.Eng.
Kp = Cj, T, = CI/CO' Td = C2/Cj. and Ph. D. degrees in electrical engineering in
1988 and 1991, respectively, both from Kyushu
Note that the PID parameters are not available by this University, Fukuoka, Japan.
method when there is no positive root for a. From May 1991 to October 1992, he was an
instructor and a research associate with the De-
partment of Mechanical Engineering, University
REFERENCES of California at Berkeley, Berkeley. He is currently a research scientist at
Omron Advanced Systems, Inc., in Santa Clara, CA. His research interests
[I] K. L. Anderson, G, L. Blankenship, and L. G. Lebow, "A rule- and publications are in the areas of system identification, signal processing,
based adaptive PID controller," in Proc. 27th IEEE Can! Decision, fault diagnosis, fuzzy logic control, neural networks, and intelligent con-
Control, 1988, pp. 564-569. trol.
[2] P. 1. Gawthrop and P. E. Nomikos, "Automatic tuning of commer-
cial PID controllers for single-loop and multiloop applications," IEEE
Control Syst. Mag., vol. 10, pp. 34-42, 1990. Masayoshi Tomizuka (M'86) was born in To-
[3] J. Gertler and H-S. Chang, "An instability indicator for expert con- kyo, Japan in 1946. He received the B.S. and M.S.
trol," IEEE Control Syst. Mag., vol. 6, pp. 14-17, 1986. degrees in mechanical engineering from Keio
[4] C. C. Hang, "The choice of controller zeros," IEEE Control Svst. University, Tokyo, Japan, in 1968 and 1970, re-
Mag., vol. 9, pp. 72-75,1989. • spectively. He received the Ph.D. degree in me-
[5] C. C. Hang, K. 1. Astrorn, and w. K. Ho, "Refinements of the chanical engineering from the Massachusetts In-
Ziegler-Nichols tuning formula," Proc. lEE, PLD., vol. 138, pp. stitute of Technology, Cambridge, MA, in 1974.
111-118,1991. He is currently a Professor with the Department
[6] T. Iwasaki and A. Morita, "Fuzzy auto-tuning for PID controller of Mechanical Engineering, University of Cali-
with model classification," in Proc. NAFlPS '90, Toronto, Canada, fornia at Berkeley. He has held visiting professor
June 6-8, 1990, pp. 90-93. positions at Keio University and at the Institute of
[7] T. Kitamori, •. A method of control system design based upon partial Industrial Science, University of Tokyo. His research interests include dig-
knowledge about controlled processes," Trans. SICE Japan. vol. 15. ital, optimal, and adaptive control with applications to mechanical systems
pp. 549-555, 1979 (in Japanese). such as robotic manipulators, machine tools, and vehicles.
[8] B. C. Kuo, Automatic Control Systems, 5th ed. Englewood Cliffs, Dr. Tomizuka served as an Associate Editor for the ASME Journal of
NJ: Prentice-Hall, 1987. Dynamic Systems, Measurement and Control and IEEE Control Systems
[9] C. C. Lee, "Fuzzy logic in control systems: Fuzzy logic controller, Magazine, and currently serves as the Technical Editor of the ASME Jour-
Part I," IEEE Trans. Syst., Man, Cybern .. vol. SMC-20, pp. 404- nal of Dynamic Systems, Measurement and Control.
418, 1990.
[10] - , "Fuzzy logic in control systems: Fuzzy logic controller, Part
II," IEEE Trans. Syst. Man, Cybern.; vol. SMC-20, pp. 419-435, Satoru lsaka (S'87-M'90) received the M.S. and
1990. Ph.D. degrees in systems science from the Uni-
[II] C. G. Nesler, "Experiences in applying adaptive control to thermal versity of California, San Diego, in 1986 and
processes in buildings," in Proc. Amer. Control Con]., Boston, MA, 1989, respectively.
1985, pp. 1535-1540. From 1987 to 1989, he was an assistant re-
[12] T. 1. Procyk and E. H. Mamdani, "A linguistic self-organizing pro- searcher with the Institute of Geophysics and
cess controller," Automatica, vol. 15, pp. 15-30, 1979. Planetary Physics, Scripps Institution of Ocean-
[13] M. Sugeno, ed., Industrial Applications of Fuzzy Control. Amster- ography, La Jolla, CA. He is currently a research
dam, The Netherlands: North-Holland, 1985. scientist at Omron Advanced Systems, Inc., in
[14] T. Takagi and M. Sugeno, "Fuzzy identification of systems and its Santa Clara, CA. His research interests include
applications to modeling and control," IEEE Trans. Svst. Man, Cy- machine learning, adaptive control, and their ap-
bern., vol. SMC-15, pp. 116-132, 1985. plications to biomedical systems and factory automation.