Professional Documents
Culture Documents
Fuzzy logic controllers have been used increasingly in industrial applications. We introduce the probabilistic controller as an
alternative to fuzzy logic controllers. The probabilistic controller is a `universal' controller with a structure closely analogous to a
popular type of fuzzy logic controller, but it is not based on fuzzy logic.
1. Introduction
Fuzzy logic controllers are currently being used in microwave ovens, train systems, robots, appliances, and
many other applications [1, 2]. According to Kosko and
Isaka [1], the Ministry of International Trade and Industry estimated that Japan produced about two billion
dollars worth of fuzzy products in 1992.
The fuzzy logic controller was originally designed, in
part, to mimic the thought process of a human operator
[3, 4]. Ebrahim Mamdani, creator of an early fuzzy logic
controller, asserted in some applications that the fuzzy
controller outperformed human operators and slightly
improved on the conventional controller ([5], p. 117).
Cheeseman has argued, however, that ``If these (controller operators) were trained in probability theory, they
might get better performance'' ([5], p. 184). In this paper
we propose a probabilistic controller that does not use
fuzzy logic, but does use similar rules for process control.
The possibility of developing such a controller has been
mentioned by Myron Tribus ([6], p. 3) and Bart Kosko
([5], p. 207). In the probabilistic approach one can choose
any probability measure that makes sense or, if one so
chooses, construct an empirical probability measure.
Thus, characterization of uncertainty by using probability is just as exible as fuzzy characterizations, but it
oers the additional advantage that the theory of probability and statistics can be invoked if one wishes.
This paper begins in Section 2 with a discussion of the
denitions used in fuzzy logic. We present an example in
Section 3 of the most basic fuzzy logic controllers. In
Section 4 the basic probabilistic controller is introduced.
In Section 5 probabilistic controllers based on two input
variables are discussed. In Section 6 we show that the
probabilistic controller is a `universal approximator'.
Equivalent performance of the probabilistic controller and
0740-817X
1997 ``IIE''
460
uncertainty for which probability theory is inadequate.
Interested readers are referred to Laviolette and Seaman
[12].
461
3.4. Defuzzication
Defuzzication must be used to translate the resulting
ouput membership function to a specic crisp action to be
taken. Several methods have also been given for defuzzication (see [3, 4, 6, 16] and [21], pp. 203209). Here we
summarize the most common method ([6], p. 309), the
center of area (gravity) method. The center of area is
given by
,Z
Z
mB zdz ;
1
COA zmB zdz
z
462
the centroid of 63.033 found by using correlation minimum encoding with the maximum operator.
The type of controller shown in our example is referred
to as a Type II controller. This seems to be the most
prevalent controller in the fuzzy literature. However, Jang
[23] categorizes all fuzzy controllers as Type I, Type II, or
Type III controllers. The Type II and Type III controllers
are also referred to as Mamdani and Sugeno controllers,
respectively. The outputs of Type I and Type III controllers are weighted averages of functions of the input
variables with the weights based on input memberships.
For more information on Type I and Type III controllers,
see Terano et al. ([24], pp. 160165) or Jang [23]. For a
general discussion of all aspects of fuzzy controllers see
Driankov et al. [21].
An important characteristic that should be noted regarding the basic FLC is that the fuzzy logic controller
does not allow variability in the control action for a given
value of the crisp input. A given crisp input always yields
the same crisp output value. However, Yager and Filev
([25], pp. 342343) have suggested an ad hoc method for
incorporating variability into the control action by scaling the output membership function to create a probability density function. A random control setting is then
generated from this density function.
4. A probabilistic controller
The probabilistic controller, discussed briey by Laviolette et al. [26], consists of a structure very similar to
the FLC. Specically, there are several classications in
which the input variable X can be placed. There is a
family of probabilities of classication in the classes indexed by X x0 . Note that we are imposing the condition
X
Pclass i; X x0 1 8 x0 :
i
Note that these probabilities could represent the probability that a human operator would classify the input as
being in a given class. The use and advantages of probabilities in place of the membership functions is discussed
and justied in detail by Hisdal [2730].
Similarly to a fuzzy controller, the probabilistic controller consists of three stages: coding, the knowledge
(rule) base, and the conict resolution. The initial stage,
which we will refer to as coding, consists of measuring the
input variable. Once the measurement x0 is obtained, the
classication probabilities are calculated. For example,
suppose we have a situation similar to that of the FLC
example given earlier. The classication probabilities are
given in Fig. 5a. This particular choice of probabilities
was made to emulate somewhat the membership functions of Kosko given in Fig. 1a. Any temperature that
would be given only one non-zero membership in the
FLC is given a probability of one in the corresponding
463
Note that for a given input value x0 , the p.d.f. of the
adjustment variable is given by
f z p1 f1 z p 2 f2 z . . . pr fr z ;
where pk P(class k; x x0 and there are r rules in the
rule base. This is true because we are making adjustment
k with probability pk . Thus, Ez and Varz are easily
obtained by using elementary rules for means and variances. For example, if two rules (rules m and n) are
activated, then the expected control action is
Ez pm lm 1 pm ln :
temperature category. Note that each probability represents the probability that a controller would choose the
temperature category given the specic temperature.
These probabilities can be determined subjectively, estimated by direct observation of an operator or a group of
operators, or by other performance-based methods commonly applied to determine memberships. The interpretation of these probabilities is simple. For example, if
the temperature is 68 F, then 40% of the time the temperature is classied as `just right' and 60% of the time it
is classied as `warm'.
The second stage, the rule base, consists of setting the
specic control rules, just as in the FLC. Using the previous example, our controller would have the same ve
ifthen rules as the FLC. The logic behind the rules is that
if P(class i; X x0 > 0 then make adjustment i with
probability P(class i; X x0 ). Again, two or more rules
may be activated.
The adjustment is a random variable that is based on
the value of the input variable. In our example, there is a
probability density function (p.d.f.) associated with each
class of motor speed. For each class j of the adjustment,
the p.d.f. is given by fj z. In our example, adjustment
groups 1, 2, 3, 4, and 5 correspond to stop, slow, medium,
fast, and blast, respectively. The p.d.f.'s for this example
are given in Fig. 5b. It is required that the area under each
p.d.f. of the control action be equal to one by denition.
This is not usually the case for the membership functions
of the FLC. For our example, the family of indexed
probabilities and the p.d.f.'s of the control action have
been set to mimic somewhat the FLC of Kosko [6].
The nal stage, conict resolution, is much easier to
apply with the probabilistic controller than with the FLC.
464
Yager and Filev ([25], pp. 342343) have recently suggested a similar method for the FLC. This method makes
the probabilistic controller non-deterministic to some
extent, and also allows the probabilistic controller to
imitate a human controller more realistically, if this is
desired. For many applications it is unlikely that a person
would always make exactly the same control action for a
given value of the crisp input x0 . There appears to be,
however, no advantage in incorporating this variability
into an automated control system unless, as Yager and
Filev ([25], pp. 342343) state, the other types of
defuzzication method lead to control actions that violate
constraints on the control action.
In this paper we do not show a probabilistic analog to
the Type I or Type III fuzzy controllers. However, if we
replace the input membership functions with a family if
indexed probability functions, the extension is straightforward. The output would simply be a weighted average
of a function of the input value(s) with the weights given
by the indexed probabilities.
x0
x0
x0
x0
is
is
is
is
xa
xa
xb
xb
and y0
and y0
and y0
and y0
Each temperature class will have ve rules. Again, multiple rules will probably be activated to some extent.
The decision making logic stage is aected by the addition of a second input. To compute the strength of a
rule, a minimum membership or product of memberships
is usually taken for the FLC. For example, suppose a
temperature has membership of 0.6 in `cold' and the
humidity has membership of 0.3 in `just right'. Thus, rule
3 has been activated to some extent. It is common to take
the minimum of 0.3 and 0.6 as the membership in `motor
speed slows'.
If another rule is also activated, we would take the
minimum input membership or take a product of memberships as above. Once the strength of each rule is determined, the process of determining the membership
function of the control action is equivalent to that for the
one-input controller. The defuzzication is completed by
using one of the methods such as the center-of-area
method discussed in Section 3.
For the probabilistic controller, there is little eect on
the approach when a second input is added. The most
profound eect is on the size of the rule base, just as with
the FLC. We now consider the previous example with the
probabilistic controller.
In coding, the temperature and humidity measurements
would be taken just as shown with the FLC. The rule base
would also be the same set of 25 ifthen rules. The difference lies in the conict resolution stage.
For each of the ve humidity classes, a family of indexed probability functions can be dened analogously to
the set shown for the temperature classes. The interpretations of these probabilities are as discussed for temperature. During coding, the categorization probabilities
are computed for temperature and humidity. Several rules
may be activated to some extent, which will require the use
of conict resolution. Independence of the classication of
the two inputs is assumed here, and for most cases this
assumption is reasonable. If the classication of temperature, for example, depended on the humidity, this could
be incorporated into the probabilistic controller.
The expression for the p.d.f. of the action function is
somewhat more complex than for the one-input case.
Note that for each rule, two categorization probabilities
temperature and humidity are computed. These probabilities are multiplied together for each action function
465
k2
Z
m1 zdz
a1
p
X
ak
k2
Z
mk zdz
z
k2
Z
m1 zdz
b1
z
p
X
k2
Z
ak
mk zdz
z
466
zm1 zdz
Z
b1
m1 zdz
z
zm1 zdz
p
X
p
X
bk
p
X
bk
zmk zdz
z
Z
bk
m1 zdz
z
Z
ak
zmk zdz
z
!Z
k1
k1
p
X
k2
ak
k2
Z
b1
k2
p
X
m1 zdz
z
p
X
bk =ck
k1
zmk zdz
z
m1 zdz
z
zm1 zdz
zm2 zdz
b1 z
b2 z
Z
Z
p
p
P
P
m1 zdz
m2 zdz
bk
bk
k1
k1
zmp zdz
bp z
Z
:
p
P
mp zdz
bk
k1
P
Note that the coecients bi = pk1 bk sum to one and
the kth ratio of integrals is the centroid of the kth output
membership function. Therefore, the output value is
simply a weighted average of the centroids of the activated output membership functions. Recall that the
output of a probabilistic controller is a weighted average
of the centroids of the activated output distributions.
Thus, if we choose
Pan equivalent rule base, categorization
probabilities bi = pk1 bk , and output density functions
such that for each output class i, the centroid of the
output density function is equal to the centroid of the
corresponding output membership function; the resulting
probabilistic controller will produce the same output
value as the additive fuzzy system for each given input
value and, therefore, the same mapping function.
j
Theorem 3 is widely applicable because most fuzzy systems
are additive ([6], p. 207). No assumptions are made concerning the shapes of the membership functions of the
FLC. See Barrett ([35], p. 44) for an example illustrating
Theorem 3.
8. Conclusions
The probabilistic controller has the same structure as the
FLC, but does not require the use of fuzzy logic. The
same type of ifthen rules are used. Thus, the probabilistic controller can solve the same type of control problem as the FLC without the mathematics of fuzzy logic,
and similar potential applications exist. The probabilistic
controller is closely analogous to the FLC with correlation product encoding, the summation operator and
centroid defuzzication and, thus, lends a theoretical
justication to this type of approach. A probabilistic
approach is not more complicated than the FLC as stated
by Kosko ([5], p. 207). Because both types of controllers
are universal approximators, it is impossible to show in a
given application that the control performance of one is,
in general, superior to that of the other.
For a discussion of fuzzy controllers from a control
theory point of view, refer to Abramovitch [36]. For other
comparisons of fuzzy and probabilistic methods, see
Cheeseman [37] or Laviolette et al. [26].
Acknowledgement
We thank the editor for a number of helpful suggestions
on an earlier version of this paper.
References
[1] Kosko, B. and Isaka, S. (1993) Fuzzy logic. Scientic American,
July, 7681.
[2] Marks, R.J., III (ed.) (1994) Fuzzy Logic Technology and Applications, IEEE Technology Update Series, New York.
[3] Berenji, H.R. (1992) Fuzzy logic controllers, in An Introduction to
Fuzzy Logic Applications in Intelligent Systems, Yager, R.R. and
Zadeh, L.A. (eds), Kluwer Academic Publishers, Boston. pp. 69
96.
[4] Pedrycz, W. (1989) Fuzzy Control and Fuzzy Systems, Research
Studies Press, Somerset.
[5] McNeill, D. and Freiberger, P. (1993) Fuzzy Logic, Simon and
Schuster, New York.
[6] Kosko, B. (1993) Fuzzy Thinking: The New Science of Fuzzy
Logic, Hyperion, New York.
[7] Zadeh, L.A. (1965) Fuzzy sets. Information and Control, no. 8,
338353.
[8] Zadeh, L.A. (1973) Outline of a new approach to the analysis of
complex systems and decision processes. IEEE Transactions on
Systems, Man, and Cybernetics, SMC-3(1), 2844.
[9] Yager, R.R. and Zadeh, L.A. (1992) An Introduction to Fuzzy
Logic Applications in Intelligent Systems, Kluwer Academic Publishers, Norwell, MA.
[10] Cox, E. (1994) The Fuzzy Systems Handbook, AP Professional,
Boston.
[11] Bonissone, P.P. (1987) Summarizing and propagating uncertain
information with triangular norms. International Journal of Approximate Reasoning, 1, 71101.
[12] Laviolette, M. and Seaman, J.W., Jr (1994) The ecacy of fuzzy
representations of uncertainty (with discussion). IEEE Transactions on Fuzzy Systems, 2(1), 442.
467
[29] Hisdal, E. (1988) Are grades of membership probabilities? Fuzzy
Sets and Systems, 25, 325348.
[30] Hisdal, E. (1988) The philosophical issues raised by fuzzy set
theory. Fuzzy Sets and Systems, 25, 340356.
[31] Wang, L.-X. (1992) Fuzzy systems are universal approximators, in
IEEE International Conference on Fuzzy Systems, IEEE, pp. 1163
1169.
[32] Lee, J. and Chae, S. (1993) Completeness of fuzzy controller
carrying a mapping f : R1 ! R1 , in IEEE International Conference
on Fuzzy Systems, IEEE, pp. 231235.
[33] Ying, H. (1994) Sucient conditions on general fuzzy systems as
functions approximators. Automatica, 30(3), 521525.
[34] Buck, R.C. (1965) Advanced Calculus, McGraw-Hill, New York.
[35] Barrett, J.D. (1995) A probabilistic alternative to fuzzy logic
controllers. Unpublished Ph.D. dissertation, Applied Statistics
Program, University of Alabama.
[36] Abramovitch, D. (1994) Some crisp thoughts on fuzzy logic, in
Proceedings of the American Control Conference, Baltimore, MD,
pp. 168172.
[37] Cheeseman, P. (1986) Probabilistic vs. fuzzy reasoning, in Uncertainty in Articial Intelligence, Kanal, L.N. and Lemmer, L.F. (eds),
Elsevier Science Publishers, Amsterdam. pp. 8593.
Biographies
William H. Woodall holds a Ph.D. degree (1980) in statistics from
Virginia Polytechnic Institute and State University. He is currently
Russell Professor of Statistics in the Department of Management Science and Statistics and Director of the Applied Statistics Program at
the University of Alabama. His research interests are in control
charting methods and in probabilistic alternatives to fuzzy methods.
He is an associate editor of IIE Transactions on Quality and Reliability
and a member of the editorial review board of Journal of Quality
Technology.
J. Douglas Barrett holds a Ph.D. in Applied Statistics from the University of Alabama. Currently, he is Assistant Professor in the Department of Economics, Finance and Qualitative Methods at the
University of North Alabama. His research interests include statistical
process control and probabilistic alternatives to fuzzy methods.