Fuzzy
Fuzzy
com/metamorworks
Tadanari Taniguchi
IT Education Center, Tokai University, Hiratsuka, Japan
Luka Eciolaza
Department of Robotics and Automation, Mondragon University, Arrasate, Spain
Michio Sugeno
Tokyo Institute of Technology, Yokohama, Japan
Abstract—More than 40 years after fuzzy logic control edge. Due to the lack of a systematic framework to study
appeared as an effective tool to deal with complex processes, Mamdani fuzzy systems, we have witnessed growing interest in
the research on fuzzy control systems has constantly evolved. fuzzy model-based approaches with Takagi-Sugeno fuzzy sys-
Mamdani fuzzy control was originally introduced as a model- tems and singleton-type fuzzy systems (also called piecewise
free control approach based on expert’s experience and knowl- multiaffine systems) over the past decades. This paper reviews
F
I. Introduction premises. PMA systems stand independently of fuzzy set and
uzzy control was initiated by Mamdani [1] in 1974 stim- logic since there is no need to use neither linguistic variables
ulated by the Zadeh’s two seminal papers on fuzzy algo- nor MFs. In the case of PMA systems, these functions in the
rithms [2] in 1968 and linguistic analysis [3] in 1973. In premises are of a triangular shape and only play roles as param-
these papers, Zadeh presented a method of system mod- eters for interpolation which are not necessarily interpreted as
eling based on fuzzy IF-THEN rules with linguistic variables. membership functions in the conventional “fuzzy” sense. How-
The first application of fuzzy logic control was performed by ever, PMA systems derived from Mamdani may keep their sta-
Mamdani and Assilian on a laboratory steam engine [4] which tus in fuzzy control and linguistic labels could be assigned to
led to a great impact on the fuzzy control research. Indeed, singletons if necessary. Note that all three types of fuzzy systems
many fuzzy control systems have been proposed since the pub- are known to have general approximation capability for any
lication of the original paper in 1975. Generally speaking, there nonlinear functions [9]. However, compared to two other types,
are three types of fuzzy systems as classified by Sugeno in [5] T-S fuzzy modeling can drastically reduce the number of fuzzy
according to the consequent parts of IF-THEN rules. First, rules, especially for high dimensional complex systems [5].
Mamdani-type fuzzy systems [4] are defined by IF-THEN During the first ten years since Mamdani’s successful applica-
rules associated with linguistic variables as tion of fuzzy logic control, researchers were faced with a lot of
criticisms from the conventional control theorists as reported in
Rule R i: IF x 1 is M i1, x 2 is M i2, f, x m is M im [5], [10], [11].The main reason is that there was no stability analy-
THEN y is H i, i = 1, 2, f, n sis available for fuzzy control at that time [12]. To answer these
criticisms, T-S fuzzy systems were newly introduced in 1985.
where H i and M ij, j = 1, 2, f, m, are fuzzy sets, n is the Since T-S fuzzy systems are associated with linear models in their
number of fuzzy rules. Second, Takagi-Sugeno (T-S) fuzzy sys- consequents, model-based stability analysis and control design can
tems [6] are with functional consequents be performed for fuzzy systems using conventional Lyapunov-
based approaches as firstly shown in [13]. T-S fuzzy modeling can
Rule R i: IF x 1 is M i1, x 2 is M i2, f, x m is M im be used to represent exactly a nonlinear system in a compact set
THEN y is f i (x 1, x 2, f, x m), i = 1, 2, f, n. of the state space, where the nonlinearities are embedded in
membership functions [14]. Therefore, it is generally impossible to
The function f i ($) is usually linear as follows: make a non-conservative stability analysis of a given nonlinear
system with T-S fuzzy model-based approaches. More clearly, T-S
f i (x 1, x 2, f, x m) = b i + a i1 x 1 + a 2i x 2 + f + a im x m, fuzzy systems belong to a class of polytopic uncertain systems
whose uncertainties are caused by the nonlinearity of the MFs. As
where the coefficients b i, a ij, with j = 1, f, m, are constants. such, T-S fuzzy model-based stability analysis always remains con-
Third, singleton-type fuzzy systems [5] are with fuzzy rules of servative [15]. PMA models were then presented to overcome this
the following form: drawback of T-S fuzzy models. As shown in [5], PMA models are
Rule R i: IF x 1 is M i1, x 2 is M i2, f, x m is M im fully parametric like linear systems, though they are approximate
models of nonlinear systems in their nature. Because of this
THEN y is b i, i = 1, 2, f, n
advantage, it is theoretically possible to derive necessary and sufficient
where b i is a singleton, i.e., a real number. A singleton-type fuzzy stability conditions for PMA systems just as in the case of linear
system is recently called piecewise multiaffine (PMA) system in systems [16]. This cannot be the case for other types of fuzzy sys-
[7] since its input-output relation with respect to an affine-in- tems. Since T-S fuzzy systems and PMA systems include linear
control system is found to be a multiaffine function. This type of systems as a special case, it is expected that any theoretical frame-
fuzzy systems is used to be called PB (piecewise bilinear, more work developed for these both types of fuzzy systems would be
precisely biaffine) systems because for a two-dimensional case, the more general than its linear counterpart.
output can be expressed as y = ax 1 + bx 2 + cx 1 x 2 + d, for some The intention of this paper is to provide a concise overview
scalars a, b, c and d, see [5], [8]. on the fuzzy control research since the pioneering works per-
Mamdani-type fuzzy systems are linguistically understand- formed by Mamdani’s group at Queen Mary College. Here, no
able since fuzzy variables are used in both the premises and the attempt is made to comprehensively review the literature which
and matrices X ! R n # n, M ! R n # n, N ! R n # n, satisfying the Exploiting this special structure of the Lyapunov function, the
following linear matrix inequalities: following theorem can be stated.
Theorem 3. [34] Given a T-S fuzzy system (5) with z / x.
Pi + X 2 0, i ! X r (10) If there exist symmetric matrices Pi ! R n # n, with i ! X r , and
matrices M ! R n # n, N ! R n # n such that
; E 1 0, i ! X r
Pz - MA i - A <i M < *
(11)
Pi + M < - NA i N + N <
Pi 2 0, i ! X r (14)
with Pz given by (9). Then, the T-S fuzzy system (5) is asymp-
; E 1 0, i ! X r (15)
-MA i - A <i M < *
totically stable.
Pi + M < - NA i N + N <
Remark 4. Differently from Theorem 1, Theorem 2 only
allows for local stability analysis of system (5) since the system with Pi having the structure in (13). Then, the T-S fuzzy system
state has to satisfy ho i (z) # z i, i ! X r . Hence, an implicit goal (5) is globally asymptotically stable.
Qi = ; E.
time-derivatives of the MFs is not required anymore for LMI- -MA i - A i< M < *
based stability analysis. Such an assumption is not always veri- Pi + M < - NA i N + N <
fied, especially in T-S fuzzy control context [35]. However, the
special structure of the matrices Pi as shown in (13) may Then, the T-S fuzzy system (5) is globally asymptotically stable.
induce some conservatism. In fact, up to now it is still hard to Remark 7. The key difference between Theorems 3 and 4
get a definitive answer on which approach leads to less conser- lies on the extra decision variable Z. Via this slack variable, the
vative stability conditions. shape information of the MFs (16), also represented by the sca-
lars b ij, can be exploited to reduce the stability conservatism in
2) Exploiting the Knowledge of the Membership Theorem 4. This shape information can be used to deal with
Functions for Stability Relaxations relatively complex MF shapes, i.e., with some minor modifica-
The MFs used to “blend” the local linear submodels of the T-S tions, multiple constraints on the MF shapes are easily handled.
fuzzy system (3) represent the nonlinearity of system (1). Howev- However, it usually requires a preliminary optimization step to
er, these MFs have been widely considered as system uncertainty, find the tightest constraint verifying (16) that could be applied
and only their convex sum property (4) has been exploited in to the MFs of the studied system.
most of the existing works based on quadratic, fuzzy and poly- Besides the two above mainstreams, research efforts have
nomial Lyapunov functions [15]. Whenever the shape of the been also devoted to reduce the stability conservatism caused
MFs and its intrinsic time-varying characteristic are not explicitly by the sufficiency of fuzzy summations, i.e., the ways how MFs
taken into account in the stability analysis, the conservativeness are dropped out to obtain a finite set of LMI conditions.
issue still remains, see further details in [29], [47]–[49]. Although numerous results have been proposed to deal with
Several approaches have been proposed to consider explicitly this source of conservativeness [55]–[61], the most prominent
the shape of the MFs in the stability analysis. These approaches approach relies on Pólya’s theorem. Based on checking the
can be classified into two following categories [50]. First, the positiveness of multidimensional matrices, the authors in [58]
membership-function-approximation approaches exploit the applied Pólya’s theorem to derive asymptotically necessary and
MF information via alternative similar functions such as staircase sufficient LMI-based conditions for the stability and perfor-
MFs [51], piecewise linear MFs [52]. Second, the membership- mance of T-S fuzzy systems. It should be stressed that as the
bound-dependent approaches exploit the bound information of homogeneous degree of a multiple summation fuzzy Lyapu-
MFs for stability analysis [48]. In addition, the MFs image space nov function increases, the conservatism of the stability con-
and the order relations among the MFs have been also exploit- ditions decreases thanks to the introduction of more degrees
ed, see for instance [50], [53], [54]. Note that membership- of freedom. However, Pólya’s theorem based approaches are
function-approximation and image-space approaches generally conceptual rather than implementable since the computa-
lead to a higher number of convex stability constraints. Con- tional burden swiftly increases in a way that most numerical
cerning membership-bound-dependent approaches, slack vari- solvers crash.
ables are usually introduced into the stability conditions for
relaxation purposes. As an illustrative example, the following C. Takagi-Sugeno Fuzzy Control Design
theorem presents the idea in [48] on using the MFs shape infor- An important application of stability theorems is to design sta-
mation to reduce further the stability conservatism of Theorem bilizing fuzzy controllers based on T-S fuzzy models of nonlin-
3 by introducing new relaxation variables. More discussions on ear plants. The control design of T-S fuzzy models are usually
the membership-function-dependent stability analysis can be addressed with the following steps: (1) choose a specific form
found in the recent survey [29]. of the control law, (2) find the respective closed-loop T-S fuzzy
Theorem 4. (adapted from [48] and [34]) Given a T-S representation, (3) apply a set of stability analysis conditions to
fuzzy system (5) with z / x. Consider the MFs vector the closed-loop representation, and (4) transform these condi-
h (x) = 6h 1 (x) h 2 (x) f h r (x)@< such that tions into LMI-based formulations. The challenge usually lies
on the last step since without any transformation at this step,
h (x) < Sh (x) + h (x) < w + v # 0 (16) the design conditions are expressed in terms of bilinear
matrix inequalities (BMIs) instead of LMIs, leading to
where S ! R r # r , w ! R r and v ! R are given. If there exist numerical difficulties.
symmetr ic matr ices Pi ! R n # n, i ! X r , and matr ices Different (state or output) feedback control schemes can be
Z ! R 2n # 2n, M ! R n # n, N ! R n # n such that applied to T-S fuzzy models [14], [22], [46]. The most com-
monly used control law is based on the so-called parallel dis-
Z 2 0, Pi 2 0, i ! X r (17) tributed compensation (PDC) concept, for which the fuzzy
Q i - b ii Z 1 0, i ! X r (18) controller shares the same fuzzy rules and sets as the T-S fuzzy
model. As a result, a PDC controller is obtained from a convex
Q i + Q j - (b ij + b ji) Z 1 0, i, j ! X r , j 2 i (19) blending of the linear local feedback gains and the MFs of the
He = G 1 0, i ! K Z , k ! K i)
received much attention from the (fuzzy) control community. Y1i + L <i W i L i /2 -Y1i | k + Pi Fk
Only very recently, it is demonstrated in [7] that the specific rep- Y2i -Y2i | k
resentation of PMA via parametric expressions (29) enables a (37)
Gt k = ;GkE, k ! K j, j ! K NZ ,
where Gk = g ( | k). Assume that the linearized system (A, B) of 0
(1) around x = 0 is stabilizable. This means that there exists a C i = Y1i + ^L <i W i L i + H < H h /2, i ! K Z ,
linear feedback gain H such that (A + BH ) is Hurwitz with Ct j = Yt1j + ^Lt <j W j Lt j + Ht < Ht h /2, j ! K NZ .
A =; E , B = ; E (41)
2f 2g
Remark 19. The slack matrices U q and W q, for q ! K r , are
2x x = 0 2x x = 0
introduced into the conditions of Theorems 5 and 6 through the
This control gain H can be designed in advance by any lin- S-procedure [28]. This contributes to reduce the conservatism
ear control technique to guarantee some local closed-loop since the piecewise-region feature of PMA systems can be fully
properties of (1). To stabilize the PMA system (40), let us con- exploited via the constraint matrices defined in (35).
sider the state-feedback control law of the form Remark 20. The design conditions in Theorem 6 are
expressed in terms of BMIs due to the product y <k y l involved
u (x) = / h k (x) y k + Hx (42) in (44) and (46). Note that for each vertex, there are 2 n associ-
k ! Kv
ated BMI-based conditions to be verified. Hence, the value of
n
where the control vertex at the origin must be assigned as y k, for k ! K v = {k 0}, should be imposed identical for 2 con-
y 0 = 0 and other input vertices y k, with k ! K v = {0}, have to ditions concerning the same vertex to avoid sliding modes and
be designed. chattering phenomena.
Remark 18. The control law (42) is composed of two
parts. The parametric expression part R k ! K h k (x) y k is defined in
v V. Future Perspectives of Fuzzy Control Systems
the same way as x and xo , see (40). The incorporation of the This paper provides a concise discussion on the evolution of
linear feedback part Hx in (42) is obviously crucial to guaran- fuzzy control systems. Through a selective list of references, we
tee the closed-loop stability in the case where system (1) is present the historical motivations and the current research
open-loop unstable around x = 0. Note that (42) can be easily progress of three types of fuzzy systems: Mamdani-type fuzzy