On The Choice of Membership Functions in A Mamdani-Type Fuzzy Controller

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

ON THE CHOICE OF MEMBERSHIP FUNCTIONS IN

A MAMDANI-TYPE FUZZY CONTROLLER

Bernadette Bouchon-Meunier*, Mariagrazia Dotoli**, Bruno Maione***

* LAFORIA-IBP, UPMC, Case 169


4 place Jussieu 75252 Paris Cédex 05, France
phone 33 1 44277003, fax 33 1 44277000
bouchon@laforia.ibp.fr

** DEE - POLITECNICO DI BARI


4 Via Orabona 70125, Italia
phone 39 80 5460312, fax 39 80 5460410
dotoli@poliba.it

*** DEE - POLITECNICO DI BARI


4 Via Orabona 70125, Italia
phone 39 80 5460312, fax 39 80 5460410
maione@poliba.it

KEYWORDS
Fuzzy control, membership function, completeness, elicitation, sensitivity.

ABSTRACT
It is well-known that the choice of membership functions is a key problem in the design
of a fuzzy controller. The aim of this paper is thus to give a further contribution in this
direction. In particular, the performance of the Mamdani-type fuzzy controller with piece wise
linear membership functions is discussed in details, taking into account features such as
overlapping, completeness level, sensitivity and convergence of control. A discussion on the
most common elicitation methods is also performed.
ON THE CHOICE OF MEMBERSHIP FUNCTIONS IN
A MAMDANI-TYPE FUZZY CONTROLLER
Bernadette Bouchon-Meunier, Mariagrazia Dotoli, Bruno Maione

1. INTRODUCTION
One of the long standing problems of fuzzy control consists in the high number of
choices necessary to define the fuzzy controller (tuning). If the designer does not dispose of
some general criteria to follow, then the fuzzy controller (FC) is characterised by some
arbitrariness.
The designer must choose the type of fuzzification (singleton or non singleton), the
number of membership functions, the functional forms of the membership functions (piece
wise linear, Gaussian, sigmoidal), the parameters of the membership functions (fixed or tuned
during a training procedure), the conjunction operator (t-norm, t-conorm), the implication or
inference operator (Mamdani, Larsen), the aggregation operator (t-norm, t-conorm) and the

INFERENCE ENGINE

RULE BASE

Crisp Input Crisp Output


FUZZIFIER DEFUZZIFIER

INFERENCE
Fuzzy Input Fuzzy Output

FIGURE 1. FUZZY CONTROLLER STRUCTURE


type of defuzzification(centroid, maxima, height). This demonstrates the richness of fuzzy
controllers but also the need for some guidelines for their practical design.
We shall, in the following, focus on the above issues concerning the choice of the
membership functions involved in the FC.
Moreover, we shall always refer to a Mamdani-type fuzzy controller (Fig. 1) with one or
more inputs and one output, characterised by a singleton fuzzifier. Furthermore, the controlled
system will be single input - single output (SISO).
In our controller we shall adopt a singleton fuzzifier, i. e. we shall suppose the inference
engine to be provided with a precise and specific observation given by a sensor. We shall also
suppose to know the set of the linguistic rules (rule-base).
More precisely, we shall refer to a fuzzy controller with n inputs x1, x2,...,xn and one
output y, with a rule-base of the kind:
If V1 is A11 and V2 is A12 and...and Vn is A1n then W is B1 (Rule R1)
or
...
or
If V1 is Al1 and V2 is Al2 and...and Vn is Aln then W is Bl (Rule Rl)
or
...
or
If V1 is AN1 and V2 is AN2 and...and Vn is ANn then W is BN (Rule RN)
A more general classification of the kind of rules can be found in [8], where six kinds of
rules are singled out: incomplete rules, mixed rules, fuzzy statements rules, comparative rules,
unless rules, quantifier rules.
2. DETERMINATION OF MEMBERSHIP FUNCTIONS AND THEIR
CHARACTERISTICS
The definition of the membership functions is a very delicate point in the design of the
FC, because the only restriction that a membership function has to satisfy is that its values
must be in the [0,1] range. A fuzzy set can therefore, unlike a crisp one, be represented by an
infinite number of membership functions [2]: let us think, for instance, about the possibility
of defining a discrete or continuous (depending on the fuzzy set it represents) membership
function, or a triangular, trapezoidal or Gaussian one and so forth. The fact that a fuzzy set
can be described by an infinite number of membership functions is at the same time a
weakness and a strength: uniqueness is sacrificed at the advantage of flexibility, thus making
the "adjustment" of a fuzzy model possible.
In the following we shall suppose that all the universes of discourse have already been
fixed, but for seek of completeness we should remark that their size greatly affects the
effectiveness of the controller. If the size of an input universe is too small, then an input value
out of scale would not fire any rule; if, on the contrary, the size is too large, then the system
could become saturated [12].

2.1. CARDINALITY AND NORMALISATION


A main issue in the choice of the membership functions involved in a fuzzy controller is
their number, that is the cardinality of each universe of discourse of the variables. Each
element of the universe of discourse must belong to a positive degree to at least one of these
fuzzy sets, so that every real input in the range under study will be taken into account, i. e. it
will fire at least one rule. A greater resolution is achieved with a high number of membership
functions, on the other side a lower computational complexity is achieved when this number
decreases, as with it the number of rules decreases, too [8, 9, 12]. In the following we shall
suppose to know the rule-base of the FC, and therefore to have a fixed number of membership
functions.
Another point to consider is the normalisation of the membership functions involved in
the FC. Normalising each membership function, that is having at least one input in which its
value is one, is not a general rule, it is more a practical rule: in fact, scaling a not normalised
membership function between zero and unity is always made possible by dividing it by its
largest value [8]. We shall always suppose to handle normalised membership functions.

2.2. TYPE OF MEMBERSHIP FUNCTION ASSIGNMENT


Let us now discuss upon the kind of definition of the membership functions, that can be
functional or numerical. A numerical assignment of the membership functions can be easily
made by means of a truth table and is useful when the universe of discourse is discrete. If the
membership function is functionally defined, then we have many possibilities, of which we
shall recall the most commonly used.
We can have either piece wise linear functions, triangular (three parameters) or
trapezoidal (four parameters), or smoother functions such as the symmetric Gaussian function
− (x −c )2
2
gaussian(x,[σ,c]) = e 2σ (two parameters),
the generalised Bell curve
1
bell(x,[a,b,c]) = (three parameters)
x − c 2b
1+
a
or the non symmetric sigmoidal function
1
sigmoid(x,[a,c]) = −a (x− c ) (two parameters).
1+e
Other curves may be obtained by the combination of the above functions, by using
some splines curves or some inverse tangent curves.
Another characteristic to be considered is the kind of subdivision of the universe of
discourse that the membership functions realise. This is usually uniform, but in some case the
membership functions can accumulate in zones where a more accurate control, i. e. a higher
sensitivity of the controller, is desired, since most of the active rules operate in that area of the
universe of discourse.
2.3. OVERLAPPING
We can also consider the possibility of having some overlapping: that membership
functions can be made to overlap [8] is one of the great strengths of the FC, as it gives a
certain robustness to the controller, because for each possible observation at least one rule is
completely fired (i. e. it is activated with a degree of membership equal to one).
Another important characteristic of the FC membership functions, related to
overlapping, is their completeness [14].
The completeness ε of a set T of fuzzy sets that describes the characteristics of a
reference set X is defined as the minimal level ε for which the ε -cuts or ε -level sets of all
the fuzzy sets cover this set X. For example, in Fig. 2 completeness is equal to 0.5.
Completeness ε describes the attitude of the fuzzy control algorithm to infer a control
action with confidence not less then ε . When this parameter decreases, the partition of the
universe of discourse decreases in fuzziness.

1
fA fB fC fD fE fF fG
0.5
completeness
ε =0.5 y

universe of discourse
FIGURE 2. COMPLETENESS OF A SET OF FUZZY SETS
On the grounds of some heuristic considerations, some authors [7] suggest a minimum
completeness level ε = 0.25, others [9] suggest an average completeness ε = 0.50, arguing
that, in this case, for every value of the input there is always a dominant rule, i. e. with a
membership grade for that input greater than or equal to 0.5. Generally speaking, it is clear
that when completeness decreases there are more regions in the universe of discourse
characterised by a low maximal truth degree of the rules they activate, thus creating the risk of
an inefficient control. On the other side, when completeness increases, there are zones
characterised by some useless, if not harmful, redundancy.
Mizumoto [9] showed that, for a first order system with a time constant T (equal to
twenty seconds) and a dead time Td (equal to two seconds), when we have a variation of the
width W of the fuzzy sets belonging to set T, the control results vary too. A low width W,
corresponding to entirely separated sets (null completeness), generally leads to bad control
results: the output does not converge to the set point and a null control is obtained, since the
problem is not solvable. On the contrary, as W increases, i. e. as the completeness level
increases, the sets overlap more and more and we obtain convergence, but the step
response overshoot becomes important.
It is therefore convenient to use fuzzy sets of significant but not too high width, in such
a way that they can be fuzzily separated, that is with a limited overlapping.
Let us remark that what Mizumoto suggested in [9] can be easily applied to the concept
of completeness, as, if the middle point of the support of each fuzzy set describing an attribute
remains unchanged with a variation of W, high values of the width W correspond to high
values of completeness ε and vice versa. This leads us to conclude that, as an average width
W is needed, a completeness around ε = 0.5 is probably the best from the point of view of
control.

2.4. ELICITATION OF MEMBERSHIP FUNCTIONS


We have so long discussed about the characteristics that the membership functions
must have. In this section we are as well interested in how to construct the membership
functions that the FC has to use.
In [4, 11, 13] are described some methods of determining the membership functions that
are essentially based on direct methods of inquiry made on human beings and corrected using
indirect methods, through which we try to eliminate the casual and the systematic
deformations affecting the membership functions because of the low reliability of man as a
measuring device.
As a general rule, we can sort out the methods of determination of a membership
function into the following three categories [1].

2.4.1. AUTOMATIC METHODS


Such methods are used when no expert is available or in the case when there are so
many data about the case study that they can be automatically processed [3].
At present the most common methods are based on neural networks and/or on genetic
algorithms [3]. Such methods are well fitting as, just like fuzzy logic, they handle numerical
data. Besides, the incontestable advantage of these approaches lies in the fact that the experts
opinion is not necessary to define the membership function: a database made of significant
examples is sufficient. The making of the membership function develops in two phases [1]:
• 1. creation of a primary membership function which is incorrectly adjusted (say,
in some cases, aleatory);
• 2. adjustment of such primary function (optimisation) [3].
The main weakness of such automatic methods consists in their non-transparency, i. e.
in the fact that we do not have any justification of the results.

2.4.2. STATISTICAL METHODS


These methods, as well as the psychological ones, are characterised by the fact that they
make membership function with a non-empty kernel; furthermore, they describe the linguistic
attributes by gradual membership functions, realised by applying an ordering structure.
In the statistical methods some data expressed in the form of frequency histograms or
other probability curves are used as a base to construct a membership function. There is a
variety of possible conversion methods, each with its own mathematical and methodological
strengths and weaknesses [1, 6]; yet, it should always be remembered that membership
functions are not probability densities, but possibility distributions [13]. Among the most used
statistical methods there are the "Yes-No" and the "Set estimate" ones, the latter proposed by
Wang [1]. These methods are not suffering from non-transparency such as the methods listed
in the previous section.
2.4.3. PSYCHOLOGICAL METHODS
This kind of strategy is often referred to as natural extraction of membership
functionsAs a general rule, in these methods some experts in the field of application of the
case study are questioned, to make them draw or specify otherwise the membership curve
appropriate to the given problem, or else they choose a curve from an assigned set of possible
membership functions. As there is an infinite array of possible membership functions fitting a
given problem, the choice is often restricted to some predefined membership functions, for
instance simple triangular shapes with a support made of a closed interval and a kernel made
of a singleton. This simplifies the choice: the expert just has to choose the central value and
the curve slope on either side. This approach is an application of the use of predefined and
parametrized membership functions [5]. Furthermore, all these methods share the idea of a
fuzzy set made of an exact part X 0 ∪ X 1 , where X 0 = X − Supp(A ) and X 1 = Ker(A ) , and
of a gradual one X f = Supp (A ) − Ker(A ) . The most commonly used psychological methods
are the so called elementary method, the structural quantification method, the Norwich and
Turksen direct method, the so called probability method and the α-cuts method [1]. When
several experts are questioned a consensual curve is built by means of aggregation [3].

3. FURTHER INSIGHTS
In spite of the existence of all these methods, the problem of translating fuzzy logic
linguistic terms to membership functions still remains partially solved. The proliferation of
the many methods listed before clearly shows that none of them is completely satisfactory.
Apart from these general methods of defining a membership function, there is a method
described in [10] that can be used when, having already obtained a membership function with
one of the methods listed from 1. to 3., it remains defined only over some elements of the
universe of discourse, while we want it to be defined over all the set. In other words, we
suppose to have a membership function f :X → [0,1] defined only over some points
x 1 <...< x n of the definition set. Therefore we know some values f (x i ) ∈ [0,1] and some
intervals in which f ( x ) = 0 and f ( x ) = 1. For instance, in the description of the membership
function associated to the idea "small", besides the values f (x i ) ∈ ]0,1[ , there are some values
of x that are definitely big (f ( x ) = 0 ) or, on the contrary, small (f ( x ) = 1).
In short, the membership function is defined only for some points. Some interpolation is
therefore needed to have a definition of f on all X. How can we accomplish this?
In [10] it has been remarked that the values f (x i ) are someway inaccurate, as they are
the result of a measure. It is therefore reasonable to choose the rest of the values of f ( x ) in
such a way that a variation of a value of x leads to the smallest change to the value of f ( x ) ,
thus adding the least possible uncertainty to the membership grades. In other words, the
membership function must be as less sensible as possible.
In [10] it has been shown that piece wise linear functions are the least sensitive in this
sense, where sensitivity is defined as
2
 df 
S( f ) = ∫ dx
 dx 
X
where we pass over the necessary properties of differentiability of f.
It is now clear why in fuzzy control membership functions are often piece wise linear:
they are the most efficient from the point of view of sensitivity.

4. CONCLUSIONS
We focused in this paper on the choice of membership functions in a Mamdani-type
fuzzy controller, with particular regard to the performances achieved when piece wise linear
membership functions are used.
A discussion on the most common elicitation methods was performed. We came to the
conclusion that, whatever method we use to determine the membership functions, it is
suitable to use some piece wise linear functions, that lead to a lower sensitivity to an error
measure made in their determination. Furthermore, the minimum completeness level required
is ε = 0.25 , which guarantees the convergence of control, and a maximum completeness
ε = 0.5, which gives a lower overshoot and a lower settling time of the time response.
In conclusion, it seems to us to have, in this session, pointed at some general results that
can help the designer in choosing the membership functions while planning the framework of
a Mamdani-type fuzzy controller. Nevertheless, much remains to be done in this direction:
many results we pointed at can be applied only to control some particular systems, whereas
many others make it possible only to reduce the choices to be done. For these reasons,
simulation still is a valuable tool for the designer and should always be made before an
effective control of the system. Much remains to be done on these subjects in the future: a
larger number of simulations could for instance help in reducing the problem to several
classes of peculiar case studies, each characterised by the controlled system.

References

[1] ALADENISE, N., "Acquisition de connaissances imparfaites auprès d'un expert:


mise en évidence d'une fonction d'appartenance", Rapport D.E.A. I.A.R.F.A.
LAFORIA, Université Paris VI, September 1994.
[2] BEZDEK, J. C., "Fuzzy models——what are they, and why?", in IEEE Transactions
on Fuzzy Systems, 1:1, Editorial, 1993.
[3] BOUCHON-MEUNIER, B., "La logique floue et ses applications", Addison-Wesley
France, 1995.
[4] DUBOIS, D., PRADE, H., "Fuzzy sets and systems, theory and applications",
Academic Press, 1980.
[5] DUBOIS, D., PRADE, H., "New results about properties and semantics of fuzzy set
theoretic operators", in P. P. Wang and S. K. Chang, "Fuzzy sets", Plenum Press,
1986.
[6] KANTROWITZ, M., HORSTKOTTE, E., JOSLYN, C., "Answers to frequently
asked questions about fuzzy logic and fuzzy expert systems", comp.ai.fuzzy,
February 1996, ftp.cs.cmu.edu:/user/ai/pubs/faqs/fuzzy/fuzzy .faq, mkant+fuzzy-
faq@cs.cmu.edu.
[7] KOSKO, B., "Neural networks and fuzzy systems", Prentice Hall, 1992.
[8] MENDEL, J., M., "Fuzzy logic systems for engineering: a tutorial", in Proceedings
of the IEEE, VOL. 83, NO 3, MARCH 1995.
[9] MIZUMOTO, M., "Fuzzy controls under various fuzzy reasoning methods", in
Information Sciences 45, 1988.
[10] NGUYEN, T. H., KREINOVICH, V., TOLBERT, D., "A measure of average
sensitivity for fuzzy logics", in International Journal of Uncertainty, Fuzzyness and
Knowledge-Based Systems, Vol. 2, No. 4, 1994.
[11] THOLE, U., ZIMMERMANN, H. J., ZYSNO, P., "On the suitability of minimum
and product operators for the intersection of fuzzy sets", in Fuzzy Sets and Systems
2, 1979.
[12] TONG-TONG, J.-R., "La logique floue", Hermès, Paris, 1995.
[13] TURKSEN, I. B., "Measurement of fuzziness: an interpretation of the axioms of
measurement", in Fuzzy Information, Knowledge Representation and Decision
Analysis, Preprints of the IFAC Symposium, Marseille, 1983, Pergamon Press, 1984.
[14] YAGER, R.R, FILEV, D. P., "Essentials of fuzzy modeling and control", John Wiley
& Sons, 1994.

You might also like