The Multiple Facets of Fuzzy Controllers:

Look-up-Tables—A Special Class of Fuzzy

Controllers

Dimitar Filev and Hao Ying

Abstract Look-up table (LUT) controllers are among the most widely utilized
control tools in engineering practice. The reasons for their popularity include
simplicity, easy to use, inexpensive hardware implementation, and strong nonlin-
earity and multimodal behaviors that can be formalized, in many cases, only by
experimentally measured data. In a previous paper, we showed that the
two-dimensional (2D) LUT controllers and one special type of two-input Mamdani
fuzzy controllers are connected in that they have the identical input-output math-
ematical relation. We also demonstrated how to represent the LUT controllers by
the fuzzy controllers. Finally, we showed how to determine the local stability of the
LUT control systems. In the present work, we extend these results to the
n-dimensional LUT controllers and the special type of the n-input Mamdani fuzzy
controllers.

Keywords Fuzzy control ⋅ Fuzzy systems ⋅ Look-up tables ⋅ Stability theory

1 Introduction

The first wave of fuzzy system applications started in the mid 70s with the work of
Mamdani and his associates [1, 2] who demonstrated that a family of fuzzy rules
could result in a control algorithm that had performance comparable to the con-
ventional industrial controllers. Almost all of the fuzzy system applications at that
time followed the mainstream fuzzy control approach-rule-based controllers with
fuzzy predicates and reasoning mechanism [3, 4], realizing nonlinear PI, PD or

D. Filev (✉)
Research & Innovation Center, Ford Motor Company, Dearborn, MI, USA
e-mail: dfilev@[Link]
H. Ying
Department of Electrical and Computer Engineering, Wayne State University,
Detroit, MI, USA
e-mail: [Link]@[Link]

© Springer International Publishing Switzerland 2017 191

J. Kacprzyk et al. (eds.), Granular, Soft and Fuzzy Approaches
for Intelligent Systems, Studies in Fuzziness and Soft Computing 344,
DOI 10.1007/978-3-319-40314-4_10

[Link]@[Link]
192 D. Filev and H. Ying

PID-like control strategies. Most of these works were focused on solving specific
control problems, e.g. climate control (Matsushita), subway control (Hitachi),
dishwasher and locomotive wheel slip control (General Electric), control of prepaint
anticorrosion process (Ford Motor Company), vehicle transmissions (Honda &
Nissan), etc. [5–9], Utilizing some of the main advantages of fuzzy control—
implementation of intuitive control strategies based on human experience, no
requirements for an explicit plant model, rapid prototyping of the control algorithm
[10]—these applications gained quick success.
Today, the synergy between fuzzy control, neural networks, evolutionary
computing, machine learning, probabilistic/possibilistic reasoning, bio-inspired
computational intelligence methodologies, and other soft-computing theory estab-
lishes the foundations of a broader control area—intelligent control. Fuzzy control
methods are also widely used in conjunction with the conventional (“hard com-
puting”) control, diagnosis, pattern recognition, signal processing, knowledge based
algorithms and systems where they are introduced within the framework of heuristic
strategies at a higher control level (supervisory control, formalization of heuristic
task and goals) or/and synergistically with control algorithms that require subjective
information, which can be difficult to formalize within the framework of conven-
tional controllers [11].
Various techniques for designing and tuning fuzzy control algorithms have been
developed to improve the robustness and tuning of the parameters and the structure
of the fuzzy control rule-base by using the similarity between the fuzzy control, PID
control, and sliding mode control [12–14]. Significant efforts have been made to
rigorously derive and study analytical structure of fuzzy controllers, i.e. the math-
ematical relationship between the input and output of a fuzzy controller. Precise
understanding of the structure is fundamentally important because it can enable one
to analyze and design fuzzy control systems more effectively with the aid of con-
ventional control theory [14–18]. The analytical structure is determined by a fuzzy
controller’s components including input fuzzy sets, output fuzzy sets, fuzzy rules,
fuzzy inference, fuzzy logic operators, and defuzzifier. Different component choices
obviously result in different analytical structures [2, 11, 16].
In the past 15 years or so, research on improving the performance of fuzzy
control algorithms, stability analysis, and systematic design of fuzzy controllers has
focused on Takagi-Sugeno method [19] and input-output models of the plant. The
Takagi-Sugeno model is a generalization of the gain-scheduling concept—instead
of linearizing at a single operating point it enables linearization in multiple vaguely
defined regions of the state space [20]. Owing to the fuzzy decomposition the
nonlinear system is represented by a polytopic nonlinear structure of coupled linear
models that has the property of a universal approximator. The polytopic repre-
sentation establishes sufficient stability conditions for the TS system using a
common Lyapunov function for a set of Lyapunov inequalities [21]. The problem
of stability and synthesis is transformed to a convex program. A systematic design
methodology that is based on solving LMIs (linear matrix inequalities) has been
developed [22]. The TS approach with its strong theoretical foundations was able to
overcome the major critics regarding the lack of analyticity of the fuzzy control and

[Link]@[Link]
The Multiple Facets of Fuzzy Controllers: Look-up-Tables… 193

made possible to address all major topics of modern control theory. However,
despite of the progress towards development of formal analytical model-based
approaches for designing fuzzy control systems, most of the practical applications
remained centered around heuristic rule-base control. It seems that this observation
only confirms the original assertion of Mamdani who introduced fuzzy logic control
in 1974 as a powerful tool to “convert heuristic control rules stated by a human
operator into an automatic control strategy” [1].
In this chapter we are focusing on a new direction of application of Mamdani
controllers that, we believe, has received little, if any, attention. We analytically
explore the relationship between one special type of Mamdani fuzzy controllers and
the multi-dimensional look-up table (LUT) controllers—one of the most widely
used practical engineering tools in industry, especially in automotive engineering—
and derive conclusions that contribute to the analysis of the look-up tables based
control systems. Our approach is inspired by the similarity between the fuzzy
controllers and look-up tables and uses the theory of fuzzy controllers to bring new
light to the look-up based engineering technique that is usually considered a low
tech and “black art” type control tool. In some sense our approach is just the
opposite to the mainstream fuzzy control literature which uses the conventional
control theory instrumentarium to explain, analyze, and further develop fuzzy
control. Our approach benefits from the great body works on fuzzy control to derive
new knowledge and to provide a new interpretation of the look-up type controllers.
LUTs are used in engineering applications as arrays of data that describe rela-
tionships between variables. In a broad sense they represent “pseudo-equations to
make up for a lack of ‘real’ equations or perhaps to replace complicated equations
with simpler ones” [23, 24]. For example, the vehicle control systems employee
thousands of LUTs that contain calibration parameters or define control actions
under different operating conditions. The most popular LUTs are the
two-dimensional (2D) tables that define the values of one dependent (output)
variable for different combinations of two independent (input) variables and even
more single dimensional LUTs. They are used as feedforward controllers or as
containers for calibrating or gain-scheduling parameters for feedback controllers.
The reasons for the popularity of the LUTs in the automotive industry are the strong
nonlinearity and multimodal behaviors in the powertrain that can be (in many cases)
formalized only by experimentally measured data under different operating con-
ditions. In addition, the LUTs are computationally effective, and can be easily
interpreted, visualized, and tuned. Two typical LUTs representing the fuel injection
time and the ignition advance at different values of the manifold absolute pressure
(MAP) and the engine speed are as follows:
The LUT output is obtained by interpolation (usually linear) in both directions,
resulting in a bilinear mapping that will be discussed in more details below. In the
following we will show that the output of the LUTs is identical to the output of a
special type of Mamdani FLCs.
The chapter is organized in two parts. The first part analyzes the relationship
between the Mamdani controllers and the LUT controllers. The main result is a
theorem proving the equivalence between one special type of Mamdani controllers

[Link]@[Link]
194 D. Filev and H. Ying

with m input variables and the m-dimensional LUT. This result provides the
framework for describing and analyzing the LUT as fuzzy controllers. In the second
part we show how to determine the local stability of feedback control systems
involving the m-dimensional LUT controllers. These findings extend our previous
results on the 2D cases [25].

2 On the Relationship Between the Mamdani FLCs

and the LUT Controllers

2.1 A Special Type of Mamdani FLCs

The general class of Mamdani FLCs is a rule-base type system that has m input
variables, designated as xi ðnÞ, i = 1, 2, …, m, where n signifies sampling instance.
xi ðnÞ may be a state variable or an input variable computed using the current and/or
historical output of a dynamic plant to be controlled (e.g., yðnÞ and yðn − 1Þ) as well
as target output signal SðnÞ. This means the input space to be m-dimensional. xi ðnÞ
is multiplied by a scaling factor ki , resulting in the scaled input variable. For
simplicity, we will use xi ðnÞ to represent the scaled variable. This will not cause
confusion because only the scaled input variables will be needed in the rest of the
chapter. The universe of discourse for xi ðnÞ is partitioned into Mi intervals. Like
most FLCs in the literature, each interval has at least one fuzzy set defined over it.
The j-th fuzzy set of xi ðnÞ is designated as Ai,̃ j whose membership function is
denoted μAi,̃ j ðxi Þ. Ai,̃ j can be any types. The fuzzy controller uses a total of
m
M = ∏ Mi fuzzy rules, each of which is in the following format:
i=1

̃ I1 AND . . . AND xm ðnÞ is Am,

IF x1 ðnÞ is A1, ̃ Im THEN uðnÞ is V k̃ ð1Þ

where the output fuzzy sets V k̃ , k = 1, …, M, cover the universe of uðnÞ. The
membership functions of V k̃ are denoted μV ̃k ðuÞ and are limited to the singleton
type. That is, V k̃ is nonzero only at one location in the universe of discount for uðnÞ
and the nonzero value is designated as Vk . The fuzzy AND operator is the product
operator

M
τh ðxÞ = ∏ μÃj, I ðxj Þ ð2Þ
j
j=1

where x = ½x1 ðnÞ⋯⋯xm ðnÞ to define the degrees of firing the rules. As for rea-
soning, any fuzzy inference method may be used in the rules. It will produce the
same inference outcome because the output fuzzy sets are of the fuzzy singleton
type (we’ll limit the discussion to the case of fuzzy singleton; the extension to fuzzy

[Link]@[Link]
The Multiple Facets of Fuzzy Controllers: Look-up-Tables… 195

sets of general shape can be found in [14]). The popular centroid defuzzifier is
employed to combine the inference outcomes of the individual rules:

M
∑ τh ðxÞ ⋅ Vh
h=1
uðnÞ = M
ð3Þ
∑ τh ðxÞ
h=1

Here, τh ðxÞ is the resulting membership of executing all the fuzzy logic AND
operations in the h-th rule whereas Vh signifies the nonzero value of the singleton
output fuzzy set in the rule.
The membership functions of the input fuzzy sets can be of general shape. The
only constraint on their selection is to guarantee a complete coverage of the
Cartesian product space of all the input variables. Expression (3) defines a deter-
ministic mapping between the inputs and the output of the Mamdani FLCs. For
finite universes of the inputs and output, which is always the case for real-world
applications, the mapping can be approximated by a LUT. For a predefined rule
base and membership functions the LUT can be calculated in advance as part of the
FLC design process. The output of the FLC can be inferred from the LUT by
interpolation. This simple LUT realization can be applied to any type of controller
but it is especially effective in the case of Mamdani FLC because it eliminates the
tedious calculations of the degrees of firing using (2)—an operation that might
require significant computational resource and time. Almost all references regarding
the LUT in the literature on fuzzy control followed the pattern described above—
the LUTs were considered as implementation tools approximating the FLCs and
their properties have not been analyzed or discussed in the framework of fuzzy
control.
In the following we’ll show that under certain assumptions the LUTs can be
identical to the FLCs, and that implementation of specific FLCs can be computa-
tionally effective and simple, comparable to the implementation of the PID con-
trollers. We’ll also show that the equivalence between the FLC and LUT can be
used to introduce a systematic approach to the local stability analysis of LUT
controllers.
In order to simplify the notations we’ll first limit the discussion to the 2D case,
i.e., assuming two input variables x1 ðnÞ and x2 ðnÞ with the corresponding universes
partitioned into intervals covered by fuzzy sets A e1, n1 and A
e1, 1 , . . . , A e2, 1 , . . . , A
e1, n2
since this type of FLCs covers the most common cases of PI- and PD-like FLCs.
Results will be further generalized to multiple input variables.
The maximal number of rules that are determined by this partitioning is
n = n1 × n2 :

[Link]@[Link]
196 D. Filev and H. Ying

̃ 1 AND x2 ðnÞ is A2,

IF x1 ðnÞ is A1, ̃ 1 THEN uðnÞ is V11
̃ 1 AND x2 ðnÞ is A2,
IF x1 ðnÞ is A1, ̃ 2 THEN uðnÞ is V12
ð4Þ
...
̃ n1 AND x2 ðnÞ is A2,
IF x1 ðnÞ is A1, ̃ n2 THEN uðnÞ is Vn1 n2

In this work we’ll make one additional assumption on the type of the fuzzy sets
e e1, n1 and A
A1, 1 , . . . , A e2, 1 , . . . , A
e1, n2 —they are defined by the normal triangular
membership functions. “Normal” means for any values of x1 ðnÞ and x2 ðnÞ the
corresponding membership values of the two neighboring fuzzy sets sum to one
(Fig. 1 and Table 1). That is

μAi,̃ j ðxi Þ + μAi,̃ j + 1 ðxi Þ = 1, j = f1, . . . , ni − 1g, i = f1, 2g

These membership functions are analogous to the concept of B-splines [26]. One
can see from Fig. 1 that the normality assumption implies that for any input value
x1 ðnÞ(or x2 ðnÞ) at least one but no more than two of the corresponding membership
grades A1, ̃ s and Ã1, s + 1 (or Ã2, t and Ã2, t + 1 ) are nonzero and consequently at least
one but no more than four of the fuzzy rules (4), including Ã1, s , Ã1, s + 1 , Ã2, t , and
̃ t + 1 , will have nonzero degrees of firing. The normality assumption also means
A2,
that the membership functions of the input variables are uniquely defined by the

Fig. 1 Normalized membership function of the fuzzy sets of input variables x1 ðnÞ and x2 ðnÞ that
satisfy the normality requirement

[Link]@[Link]
The Multiple Facets of Fuzzy Controllers: Look-up-Tables… 197

Table 1 Sample LUTs Speed (RPM)

representing fuel injection
MAP (kPA) 500 1000 1500
time and ignition advance for
different engine operating 50 4.02 4.06 4.09
conditions—manifold 40 3.36 3.39 3.36
absolute pressure (MAP) and 30 3.24 3.27 3.31
engine speed [31]
20 3.04 2.89 3.04
Speed (RPM)
MAP (kPA) 500 1000 1500
50 10 16 18
40 9.4 15.7 17.3
30 8.6 15 16.8
20 7.9 14.5 15.9

universe parameters a1, s and a2, t , s = f1, . . . , n1 − 1g, t = f1, . . . , n2 − 1g, matching
their maxima.

2.2 Relationship Between the Mamdani FLCs and the LUT

Controllers

Because of the popularity of 2D LUT controllers in real-world applications, we will

first show that if a 2D LUT that has as entries the singleton values Vs, t , and with
rows and columns defined by the universe parameters a1, s and a2, t (Table 2) then
the LUT output will be identical to the above fuzzy controllers. In other words, a
two-input Mamdani FLC with the normal triangular membership functions of the
inputs, singleton consequents and the product AND operator is identical to a 2D
LUT that has rows, columns and entries that are defined by the corresponding
parameters of the membership functions and the singleton consequents of the FLC.

Table 2 LUT with rows and columns defined by the universe parameters a1, s , a2, t , and entries—
the singleton consequents Vs, t , s = f1, . . . , n1 − 1g, t = f1, . . . , n2 − 1g
x1
x2 a2, n2 Vn1 , n2
a2, n2 − 1
…
a2, t+1 Vs , t+1 Vs+1, t+1
a2, t Vs , t Vs+1, t
…
a2,2
a2,1 V1,1 V1,2
a1,1 a1,2 … a1, s a1, s+1 … a1, n1 − 1 a1, n1

[Link]@[Link]
198 D. Filev and H. Ying

According to Fig. 2 for a set of arbitrarily chosen x1 ðnÞ and x2 ðnÞ we get for the
degrees of firing of the affected rules:

a1, s + 1 − x1 ðnÞ a2, t + 1 − x2 ðnÞ

τs, t ðxÞ = μÃ1, s ðx1 Þ ⋅ μA2,̃ t ðx2 Þ = ×
a1, s + 1 − a1, s a2, t + 1 − a2, t
a1, s + 1 − x1 ðnÞ x2 ðnÞ − a2, t
τs, t + 1 ðxÞ = μÃ1, s ðx1 Þ ⋅ μA2,̃ t + 1 ðx2 Þ = ×
a1, s + 1 − a1, s a2, t + 1 − a2, t
ð5Þ
x1 ðnÞ − a1, s a2, t + 1 − x2 ðnÞ
τs + 1, t ðxÞ = μÃ1, s + 1 ðx1 Þ ⋅ μA2,̃ t ðx2 Þ = ×
a1, s + 1 − a1, s a2, t + 1 − a2, t
x1 ðnÞ − a1, s x2 ðnÞ − a2, t
τs + 1, t + 1 ðxÞ = μÃ1, s + 1 ðx1 Þ ⋅ μA2,̃ t + 1 ðx2 Þ = ×
a1, s + 1 − a1, s a2, t + 1 − a2, t

The four firing levels are positive and sum to one. Therefore, the defuzzifier (3)
makes the output of the FLC uðnÞ to interpolate between the corresponding sin-
gletons according to the current values of the degrees of firing as functions of the
current input values x1 ðnÞ and x2 ðnÞ:

uðnÞ = τs, t ðxÞVs, t + τs, t + 1 ðxÞVs, t + 1 + τs + 1, t ðxÞVs + 1, t + τs + 1, t + 1 ðxÞVs + 1, t + 1 ð6Þ

Alternatively, from the grid representation of the LUT in Table 1 (Fig. 2), we
can obtain uðnÞ by interpolating first between Vs, t and Vs + 1, t , and between Vs, t + 1
and Vs + 1, t + 1 along x1 ðnÞ axis. The intermediate interpolated values vt and vt + 1 (the
order of interpolation does not matter) are

a1, s + 1 − x1 ðnÞ x1 ðnÞ − a1, s

vt = Vs, t + Vs + 1, t ð7Þ
a1, s + 1 − a1, s a1, s + 1 − a1, s

Fig. 2 Grid representation of

the LUT of Table 2

[Link]@[Link]
The Multiple Facets of Fuzzy Controllers: Look-up-Tables… 199

a1, s + 1 − x1 ðnÞ x1 ðnÞ − a1, s

vt + 1 = Vs, t + 1 + Vs + 1, t + 1 ð8Þ
a1, s + 1 − a1, s a1, s + 1 − a1, s

Similarly, by interpolating between vt and vt + 1 along x2 ðnÞ axis and substituting

for vt and vt + 1 according to (8) and (9) we get for the interpolated value uT ðnÞ that
is inferred by the LUT:

a2, t + 1 − x2 ðnÞ x2 ðnÞ − a2, t

uT ðnÞ = vt + vt + 1
a2, t + 1 − a2, t a2, t + 1 − a2, t
 
a2, t + 1 − x2 ðnÞ a1, s + 1 − x1 ðnÞ x1 ðnÞ − a1, s
= Vs, t + Vs + 1, t
a2, t + 1 − a2, t a1, s + 1 − a1, s a1, s + 1 − a1, s
 
x2 ðnÞ − a2, t a1, s + 1 − x1 ðnÞ x1 ðnÞ − a1, s
+ Vs, t + 1 + Vs + 1, t + 1
a2, t + 1 − a2, t a1, s + 1 − a1, s a1, s + 1 − a1, s
ð9Þ
a2, t + 1 − x2 ðnÞ a1, s + 1 − x1 ðnÞ a2, t + 1 − x2 ðnÞ
= Vs, t +
a2, t + 1 − a2, t a1, s + 1 − a1, s a2, t + 1 − a2, t
x1 ðnÞ − a1, s x2 ðnÞ − a2, t a1, s + 1 − x1 ðnÞ
Vs + 1, t + Vs, t + 1
a1, s + 1 − a1, s a2, t + 1 − a2, t a1, s + 1 − a1, s
x2 ðnÞ − a2, t x1 ðnÞ − a1, s
+ Vs + 1, t + 1 .
a2, t + 1 − a2, t a1, s + 1 − a1, s

By comparing with expression (6) we finally obtain:

uT ðnÞ = τs, t ðxÞVs, t + τs, t + 1 ðxÞVs, t + 1 + τs + 1, t ðxÞVs + 1, t + τs + 1, t + 1 ðxÞVs + 1, t + 1 = uðnÞ

Therefore, the FLC and the LUT produce the same results.
The equivalence between the FLC and the LUT suggests a comprehensive
representation and effective computational realization of the FLC with the normal
triangular membership functions. We’ll illustrate the opportunity for simplifying the
implementation of the FLC based on its equivalence to the LUT on the showcase of
a simple PI/PD-like FLC. This FLC uses the rule base (5) and has two inputs—the
scaled error between the set point and the plant output, EðnÞ = Ke eðnÞ and the scaled
difference of the error, ΔEðnÞ = KΔe ΔeðnÞ = KΔe ðeðnÞ − eðn − 1ÞÞ, and one output
that coincides with the control variable uðnÞ (i.e., PD-like FLC) or its difference
ΔuðnÞ = uðnÞ − uðn − 1Þ (i.e., PI-like FLC) [14]. A prototypical set of rules of this
FLC can be derived from the meta-rules defining a common sense control strategy:
• If EðnÞ and ΔEðnÞ are zero, then maintain present control output
• If EðnÞ is tending to zero at a satisfactory rate, then maintain present control
output
• If EðnÞ is not self-correcting, then a nonzero ΔuðnÞ is added to present control
output, depending on the sign and magnitude of EðnÞ and ΔEðnÞ

[Link]@[Link]
200 D. Filev and H. Ying

Table 3 Example of a rule E(n) →

base of the Mamdani FLC
Ne Ze Pe
ΔEðnÞ Nd N* NM* Z*
↓ Zd N* Z* P*
Pd Z* PM* P*

Table 3 shows an example of a PI-like Mamdani FLC rule base that is based on
partitioning of the universes of EðnÞ and ΔEðnÞ into three fuzzy sets—Negative
(Ne), Zero (Ze), and Positive (Pe) for EðnÞ, and Negative (Nd), Zero (Zd), and
Positive (Pd) for ΔEðnÞ,and fuzzy singletons (real values) Negative (N* = −1),
Negative Medium (NM* = −0.4), Zero (Z* = 0), Positive Medium (PM* = 0.4),
and Positive (P* = 1) that are defined on the ΔuðnÞ universe:
One can easily transfer the rule base in Table 3 to the generic rule base format
(4) by letting

A1,̃ 1 = Ne ; A2,
̃ 1 = Nd ; V ̃1, 1 = N *
̃ 2 = Ne ; A2,
A1, ̃ 2 = Zd ; V ̃2, 2 = N *
̃ 3 = Ne ; A2,
A1, ̃ 3 = Pd ; V ̃3, 3 = Z *
̃ 4 = Ze ;
A1, Ã2, 4 = Nd ; V ̃4, 4 = N *
̃ 5 = Ze ;
A1, Ã2, 5 = Zd ; V ̃5, 5 = Z *
̃ 6 = Ze ; Ã2, 6 = Pd ; V ̃6, 6 = P*
A1,
̃ 7 = Pe ; A2,
A1, ̃ 7 = Nd ; V ̃7, 7 = Z *
̃ 8 = Pe ;
A1, ̃ 8 = Zd ; V ̃8, 8 = P*
A2,
̃ 9 = Pe ;
A1, ̃ 9 = Pd ; V ̃9, 9 = P*
A2,

If we consider the normal triangular membership functions of EðnÞ and ΔEðnÞ in

Fig. 3 we can replace the rule base formulation (Table 3) by the LUT in Table 4.
The output of this LUT equivalent form of the FLC can be calculated by only
one line of MATLAB code—line 11 in Fig. 5.
The equivalent LUT form of the FLC combines all the advantages of the FLC
with the efficiency and transparency of the LUT. It allows for a computational

Fig. 3 Normal triangular

membership functions of the
fuzzy sets of the error EðnÞ
and its difference ΔEðnÞ

[Link]@[Link]
The Multiple Facets of Fuzzy Controllers: Look-up-Tables… 201

% LUT FLC
% FLC Definition
% Universes of Error e and Error Difference de
E=[-1 0 1];
DE=[-2 0 2];
% Control Singletons (each row corresponds to one value of de)
U=[-1 -.4 0; -1 0 1; 0 0.4 1];
% End of FLC Definition
% FLC Output uk for given values of the erorr ek and error difference dek
uk=interp2(E,DE,U,ek,dek);

Fig. 4 Sample MATLAB implementation of the LUT equivalence of the FLC. The actual FLC is
calculated in last line

Table 4 The LUT that is equivalent to the Mamdani FLC with the rule base by Table 3 and the
membership functions by Fig. 4
E(n) →
-1 0 1
ΔEðnÞ −2 -1 −.0.4 0
↓ 0 -1 0 1
2 0 0.4 1

Fig. 5 Four sub-rectangles partitioning the rectangle ða1, s + 1 − a1, s Þ × ða2, t + 1 − a2, t Þ

efficient implementation of the FLC that does not require dealing with the rule base,
rules firing, and defuzzification, while exactly preserving all the properties of the
FLC.
From Fig. 2 and expression (9) we can see that the interpolated value of the 2D
LUT that is inferred for two arbitrary inputs x1 ðnÞ and x2 ðnÞ is essentially the
weighted average of the nodes Vs, t , Vs + 1, t , Vs, t + 1 , and Vs + 1, t + 1 over the nor-
malized areas aa1,1,ss++11−−xa11,ðnÞs . aa2,2,t +t +11−−xa2 2,ðnÞt , aa1,1,ss++11−−xa11,ðnÞs .ax2,2 ðnÞ − a2, t x1 ðnÞ − a1, s a2, t + 1 − x2 ðnÞ
, .
t + 1 − a2, t a1, s + 1 − a1, s a2, t + 1 − a2, t
, and
x1 ðnÞ − a1, s x2 ðnÞ − a2, t
a1, s + 1 − a1, s .a2, t + 1 − a2, t of the four sub-rectangles. These sub-rectangles are determined

[Link]@[Link]
202 D. Filev and H. Ying

by the partitioning of the rectangle ða1, s + 1 − a1, s Þ × ða2, t + 1 − a2, t Þ by x1 ðnÞ and
x2 ðnÞ, and each of them is associated with one of four nodes Vs, t , Vs + 1, t , Vs, t + 1 , and
Vs + 1, t + 1 —see Fig. 6 that illustrates this partitioning. The areas of the rectangles are
used as a measure of closeness between the input ðx1 ðnÞ, x2 ðnÞÞ and the nodes in
the 2D space. It is easy to see that the weights are positive and sum to one.
In the above discussion we demonstrated the equivalence between the FLCs
and the LUT considering the commonly used two-inputs, single-output model of
a FLC and the 2D LUT that is widely accepted in industry. We now extend this
result to cover m-dimensional LUT controllers, which is important both theo-
retically and practically. That is, we’ll use the above observation to derive the
expression for the output inferred by an mD LUT and to prove its equivalence
to an mD FLC.
Theorem 1 A Mamdani FLC with m inputs partitioned into (ni + 1), i = {1, 2, …,
m}, fuzzy subsets with the normal triangular membership functions, product AND
aggregation of the input fuzzy subsets, and singleton consequents is equivalent to
an m-dimensional LUT with grid points defined by the arguments of the maxima of
the membership functions of the input variables and grid point entries corre-
sponding to the consequents.
Proof Assume an mD LUT where the inputs are divided into an m dimensional
grid of n1, n2, …, nm of m dimensional cells defined by the ordered grid points
ai, 1 , ai, 2 , . . . , ai, si , ai, ni + 1 , i = f1, 2, . . . , mg and corresponding functional values
(LUT entries) V1, 1, ..., 1 , V1, 1, ..., 2 ,. . . , V1, 1, ..., si , V1, 1, ..., si + 1, . . . , V1, 1, ..., nm + 1 ,…,
Vn1 + 1, n2 + 1, ..., nm + 1 where si = f1, 2, . . . , ni g, i = f1, 2, . . . , mg are arbitrary inter-
mediate points. Assume now an mD FLC, i.e. an m-input, single output fuzzy
system with inputs partitioned into (n1 + 1), (n2 + 1), …, (nm + 1) normal fuzzy
subsets and singleton consequents (Fig. 6). The LUT grid points and entries
coincide with the corresponding maxima of the membership functions and the
consequents of the FLC. This type of fuzzy partitioning defines a family of (n1 + 1)
(n2 + 1) … (nm + 1) rules with fuzzy predicates:

Fig. 6 Membership functions of the fuzzy sets of the i-th FLC input xi

[Link]@[Link]
The Multiple Facets of Fuzzy Controllers: Look-up-Tables… 203

IFx1 isA1, 1 AND x2 is A2, 1 AND . . . xm is Am, 1 THEN u is V1, 1, ..., 1

IF x1 is A1, 1 AND x2 is A2, 1 AND . . . xm is Am, 2 THEN u is V1, 1, ..., 2
...
IF x1 is A1, 1 AND x2 is A2, 1 AND . . . xm is Am, sm THEN u is V1, 1, ..., sm
IF x1 is A1, 1 AND x2 is A2, 1 AND . . . xm is Am, sm + 1 THEN u is V1, 1, ..., sm + 1
...
IF x1 is A1, 1 AND x2 is A2, 1 AND . . . xm is Am, nm + 1 THEN u is V1, 1, ..., nm + 1
...
IF x1 is A1, n1 + 1 AND x2 is A2, n2 + 1 AND . . . xm is Am, nm + 1 THEN u is Vn1 + 1, n2 + 1, ..., nm + 1

Consider the m dimensional LUT cell Cs formed by a1, s1 ≤ x1 ðnÞ ≤ a1, s1 + 1 ,

a2, s2 ≤ x2 ðnÞ ≤ a2, s2 + 1 , am, sm ≤ xm ðnÞ ≤ am, sm + 1 , and si = f1, 2, . . . , ni g, i = f1, 2,
. . . , mg. The cell is defined by the hyperrectangle

Cs ≡ ða1, s1 + 1 − a1, s1 Þ × ða2, s2 + 1 − a2, s2 Þ × . . . × ðam, sm + 1 − am, sm Þ.

The hyperrectangle Cs has 2m nodes, Vs1 , s2 , ..., sm , Vs1 , s2 , ..., sm + 1 , . . . ,

Vs1 + 1, s2 + 1, ..., sm + 1 , which are associated with the LUT entries. The hyperrectangle
Cs is divided by the inputs x1 ðnÞ, x2 ðnÞ, . . . , and xm ðnÞ into 2m sub-
hyperrectangles:

Cs1 + 1, s2 + 1, ..., sm + 1 ≡ ða1, s1 + 1 − x1 ðnÞÞ × ða2, s2 + 1 − x2 ðnÞÞ × . . . × ðam, sm + 1 − xm ðnÞÞ

Cs1 + 1, s2 + 1, ..., sm ≡ ða1, s1 + 1 − x1 ðnÞÞ × ða2, s2 + 1 − x2 ðnÞÞ × . . . × ðxm ðnÞ − am, sm Þ
Cs1 , s2 , ..., sm ≡ ðx1 ðnÞ − a1, s1 Þ × ðx2 ðnÞ − a2, s2 Þ × . . . × ðxm ðnÞ − am, sm Þ

Each of the sub-hyperrectangles includes one of the nodes of the cell Cs . Using
the definition of a volume of a hyperrectangle as a product of its sides [27] (note
that this definition of the volume of a hyperrectangle is consistent with the defi-
nitions of the area of a rectangle in 2D and the volume of a cuboid in the 3D space)
we can express the volume of the cell Cs as a sum of the volumes of the
sub-hyperrectangles Cs1 , s2 , ..., sm , Cs1 , s2 , ..., sm + 1 , . . . , Cs1 + 1, s2 + 1, ..., sm + 1 :

ða1, s1 + 1 − a1, s1 Þða2, s2 + 1 − a2, s2 Þ . . . ðam, sm + 1 − am, sm Þ = ða1, s1 + 1 − x1 ðnÞÞ

ða2, s2 + 1 − x2 ðnÞÞ . . . ðam, sm + 1 − xm ðnÞÞ +
ða1, s1 + 1 − x1 ðnÞÞða2, s2 + 1 − x2 ðnÞÞ . . . ðxm ðnÞ − am, sm Þ + . . .
+ ðx1 ðnÞ − a1, s1 Þðx2 ðnÞ − a2, s2 Þ . . . ðxm ðnÞ − am, sm Þ

Similarly to the 2D case, the volumes of the sub-hyperrectangulars reflect the

closeness of the input ðx1 ðnÞ, x2 ðnÞ, . . . xm ðnÞÞ to the nodes of Cs . By associating
the normalized volumes of the sub-hyperrectangles with the corresponding nodes of

[Link]@[Link]
204 D. Filev and H. Ying

Cs − Vs1 , s2 , ..., sm Vs1 , s2 , ..., sm + 1 , . . . , Vs1 + 1, s2 + 1, ..., sm + 1 —we get the interpolated value
that is inferred by the rD LUT:

1
uT ðnÞ = ðða1, s1 + 1 − x1 ðnÞÞða2, s2 + 1 − x2 ðnÞÞ . . . ðar, sr + 1 − xr ðnÞÞ
K
Vs1 + 1, s2 + 1, ..., sm + 1 + ða1, s1 + 1 − x1 ðnÞÞða2, s2 + 1 − x2 ðnÞÞ ð10Þ
ðxr ðnÞ − ar, sr ÞðVs1 + 1, s2 + 1, ..., sm Þ + . . . + ðx1 ðnÞ − a1, s1 Þ
ðx2 ðnÞ − a2, s2 Þ . . . ðxr ðnÞ − ar, sr ÞVs1 , s2 , ..., sm Þ

where the normalizing factor K = ða1, s1 + 1 − a1, s1 Þða2, s2 + 1 − a2, s2 Þ . . .

ðam, sm + 1 − am, sm Þ is the volume of the hyperrectangle ða1, s1 + 1 − a1, s1 Þ × ða2, s2 + 1
− a2, s2 Þ × . . . × ðam, sm + 1 − am, sm Þ. It is easy to see that by combining pair of the
terms with all but one elements identical, e.g. the first two terms in (), the weights
will sum to one.
Consider now that the same arbitrary input values x1 ðnÞ, x2 ðnÞ, and xm ðnÞ where

a1, s1 ≤ x1 ðnÞ ≤ a1, s1 + 1 , a2, s2 ≤ x2 ðnÞ ≤ a2, s2 + 1 , . . . , am, sm ≤ xm ðnÞ ≤ am, sm + 1 ,

and si = f1, 2, . . . , m − 1g, i = f1, 2, . . . , mg are applied to the FLC. As we can

see from Fig. 6 only the following 2m rules:

IF x1 is A1, s1 AND x2 is A2, s2 AND . . . xm is Am, sm THEN u is Vs1 , s2 , ..., sm

IF x1 is A1, s1 AND x2 is A2, s2 AND . . . xm is Am, sm + 1 THEN u is Vs1 , s2 , ..., sm + 1
...
IF x1 is A1, s1 + 1 AND x2 is A2, s2 + 1 AND . . . xm is Am, sm + 1 THEN u is Vs1 + 1, s2 + 1, ..., sm + 1

will fire because their antecedents have nonzero membership values.

Following Fig. 6 and assuming a product AND operator the degrees of firing of
those rules are:

ðx1 ðnÞ − a1, s1 Þðx2 ðnÞ − a2, s2 Þ . . . ðxm ðnÞ − am, sm Þ

τs1 , s2 , ..., sm =
ða1, s1 + 1 − a1, s1 Þða2, s2 + 1 − a2, s2 Þ . . . ðam, sm + 1 − am, sm Þ
ðx1 ðnÞ − a1, s1 Þðx2 ðnÞ − a2, s2 Þ . . . ðxm ðnÞ − am, sm + 1 Þ
τs1 , s2 , ..., sm + 1 =
ða1, s1 + 1 − a1, s1 Þða2, s2 + 1 − a2, s2 Þ . . . ðam, sm + 1 − am, sm Þ
...
ðx1 ðnÞ − a1, s1 + 1 Þðx2 ðnÞ − a2, s2 + 1 Þ . . . ðxm ðnÞ − am, sm + 1 Þ
τs1 + 1, s2 + 1, ..., sm + 1 =
ða1, s1 + 1 − a1, s1 Þða2, s2 + 1 − a2, s2 Þ . . . ðam, sm + 1 − am, sm Þ

By applying the defuzzification law (3) we obtain for u(n) an expression that is
identical to that for the output of the LUT (10):

[Link]@[Link]
The Multiple Facets of Fuzzy Controllers: Look-up-Tables… 205

ðx1 ðnÞ − a1, s1 Þðx2 ðnÞ − a2, s2 Þ . . . ðxm ðnÞ − am, sm Þ

uðnÞ = Vs , s , ..., s
ða1, s1 + 1 − a1, s1 Þða2, s2 + 1 − a2, s2 Þ . . . ðam, sm + 1 − am, sm Þ 1 2 m
ðx1 ðnÞ − a1, s1 Þðx2 ðnÞ − a2, s2 Þ . . . ðxm ðnÞ − am, sm + 1 Þ
+ Vs , s , ..., s + 1 + . . .
ða1, s1 + 1 − a1, s1 Þða2, s2 + 1 − a2, s2 Þ . . . ðam, sm + 1 − am, sm Þ 1 2 m
ðx1 ðnÞ − a1, s1 + 1 Þðx2 ðnÞ − a2, s2 + 1 Þ . . . ðxm ðnÞ − am, sm + 1 Þ
+ Vs1 + 1, s2 + 1, ..., sm + 1
ða1, s1 + 1 − a1, s1 Þða2, s2 + 1 − a2, s2 Þ . . . ðam, sm + 1 − am, sm Þ

We point out that the properties of the FLCs with the normal triangular mem-
bership functions have been studied by a number of researchers. Most of their work
have been focused on analyzing the similarity between the FLCs and the linear PI
and PD controllers, demonstrating that for special values of the consequents the
FLCs are identical to the nonlinear PI or PD controllers (e.g., [14, 16, 28]). A more
detailed recent studies of the properties of the fuzzy models with the normal
membership functions from the perspective of bilinear systems by Sugeno and
Taniguchi [29] were partially the inspiration for our work.
In the next section we’ll show the input-output relationship of the LUT con-
trollers and how to determine the local stability of the LUT control systems—a
topic the has received very little, if any, attention although a great deal of practical
control applications are based on such LUT controllers.

3 Analytical Structure and Local Stability of the LUT

Control Systems

We will first focus on the analytical structure of the fuzzy controllers with two input
variables and will then generalize the result to the m-dimensional LTU controllers.
By simplifying (9) we obtain the analytical structure of the 2D LUT controllers,
which is also the 2D FLCs, as follows:

a2, t + 1 − x2 ðnÞ a1, s + 1 − x1 ðnÞ a2, t + 1 − x2 ðnÞ

uT ðnÞ = uðnÞ = Vs, t +
a2, t + 1 − a2, t a1, s + 1 − a1, s a2, t + 1 − a2, t
x1 ðnÞ − a1, s x2 − a2, t a1, s + 1 − x1 ðnÞ
Vs + 1, t + Vs, t + 1
a1, s + 1 − a1, s a2, t + 1 − a2, t a1, s + 1 − a1, s ð11Þ
x2 − a2, t x1 ðnÞ − a1, s
+ Vs + 1, t + 1
a2, t + 1 − a2, t a1, s + 1 − a1, s
= αx1 ðnÞ + βx2 ðnÞ + γx1 ðnÞx2 ðnÞ + δ

where α, β, γ, and δ are constants:

[Link]@[Link]
206 D. Filev and H. Ying

α = ða2, t + 1 Vs + 1, t − a2, t + 1 Vs, t + a2, t Vs, t + 1 − a2, t Vs + 1, t + 1 Þ ̸K

β = ða1, s + 1 Vs, t + 1 − a1, s + 1 Vs, t + a1, s Vs + 1, t − a1, s Vs + 1, t + 1 Þ ̸K
γ = ðVs, t − Vs + 1, t − Vs, t + 1 + Vs + 1, t + 1 Þ ̸K
δ = ða1, s + 1 a2, t + 1 Vs, t + a1, s a2, t Vs + 1, t + 1 − a1, s a2, t + 1 Vs + 1, t − a1, s + 1 a2, t Vs, t + 1 Þ ̸K
K = ða2, t + 1 − a2, t Þða1, s + 1 − a1, s Þ

That is, the LUT controllers are nonlinear controllers with constant offset term
(δ).
Likewise, simplifying (10), one can attain the nonlinear analytical structure of
the rD LUT controllers, or equivalently that of the rD FLCs:
 r
 r
uT ðnÞ = uðnÞ = ∑ βd1 ...dr ∏ xi ,
di
∑ di ≤ r ð12Þ
di ≥ 0 i=1 i=1

where the values of the constants βd1 ...dr , is determined by all the constant
parameters in (10). For concise presentation, we omit their complicated general
relations with these parameters. For any specific LUT controller, especially those
involving only a handful number of input variables which is mostly the case in
practice, βd1 ....dr , can be easily determined. For instance, when there are two input
variables, based on (11), it is obvious that β10 = α, β01 = β, β11 = γ, and β00 = δ.
We now turn our attention to the local stability determination of the rD LUT
controller regulating a nonlinear dynamic system. Without loss of generality,
assume that when the system to be controlled is at the equilibrium point of our
interest, x1 ðnÞ = . . . .. = xr ðnÞ = 0. We want to study the condition for the nonlinear
rD LUT control system to be stable at least in the area around the equilibrium point.
If both the system and the LUT controller are linearizable at the equilibrium point,
then the system stability at that point can be decided by applying Lyapunov’s
linearization method [30] to the linearized LUT controller and the linearized sys-
tem. Thus, we obtain the following result:
Theorem 2 Suppose that the rD LUT controller is used to control a nonlinear
system that is linearizable at the equilibrium point. The control system is locally
stable (or unstable) at the equilibrium point if and only if the linearized system
involving the LUT controller linearized at the equilibrium point

uT ðnÞ = β10⋯0 x1 ðnÞ + β010⋯0 x2 ðnÞ + ⋯ + β0⋯0r xr ðnÞ

is strictly stable (or unstable).

Proof First, notice that because (12) is a multi-variable polynomial, it is always
linearizable at an equilibrium point no matter where it is. Now suppose that the
equilibrium point is x1 ðnÞ = . . . .. = xr ðnÞ = 0. Linearization of (12) is achieved if all
the cross-product terms of the variables are dropped because xi ðnÞ ≈ 0 for all the

[Link]@[Link]
The Multiple Facets of Fuzzy Controllers: Look-up-Tables… 207

variables and hence the higher order cross-product terms are much smaller than
xi ðnÞ. This results in

uT ðnÞ = β10⋯0 x1 ðnÞ + β010⋯0 x2 ðnÞ + ⋯ + β0⋯0r xr ðnÞ + β0⋯0

Since the constant term β0...0 does not affect the system stability, it is removed
from the stability study. The conclusion naturally follows by using Lyapunov’s
linearization method. The method states that if the nonlinear control system is
continuously differentiable at the equilibrium point and the linearized system is
strictly stable (or unstable) at the equilibrium point, then the equilibrium point is
locally stable (or unstable) for the original nonlinear system. Q.E.D The lineariz-
ability test must be met for the system to be controlled as it is the precondition for
the theorem. A test failure only means inapplicability of the theorem; it does not
imply system instability. Theorem 2 offers some practically important advantages.
First, it is a necessary and sufficient condition. Unlike sufficient conditions or
necessary conditions, it is not conservative and is the “tightest” possible stability
condition. Second, the theorem can be used not only when the system model is
available, but also when it is unavailable but is known linearizable at the equilib-
rium point (most physical systems are linearizable).

4 Conclusion

In this chapter we expanded our previous work on the relationship between the
LUT controllers and one special type of fuzzy controller. We showed that the
multidimensional LUT’s are closely related to the fuzzy models and can be con-
sidered and analyzed, in a broad sense, a special class of fuzzy models.

References

1. Mamdani, E., Assilian, S.: An experiment in linguistic synthesis with a fuzzy logic controller.
Int. J. Man Mach. Stud. 7(1), 1–13 (1975)
2. Kickert, W., Mamdani, E.: Analysis of a fuzzy logic controller. Fuzzy Sets Syst. 12, 29–44
(1978)
3. Reznik, Leon: Fuzzy Controllers Handbook: How to Design Them, How They Work. Newnes,
Oxford (1997)
4. Zadeh, L.: Outline of a new approach to the analysis of complex systems and decision
processes. IEEE Trans. Syst. Man Cybern. 3, 28–44 (1973)
5. Sugeno, M. (ed.): “Industrial Applications of Fuzzy Logic Control.”. North Holland,
Amsterdam (1985)
6. Schwartz, D.G., Klir, G.J.: Fuzzy logic flowers in Japan. IEEE Spectr. 29, 32–35 (1991)
7. Schwartz, D.G., Klir, G.J., Lewis III, H.W., Ezawa, Y.: Applications of fuzzy sets and
approximate reasoning. Proc. IEEE 82, 482–498 (1994)

[Link]@[Link]
208 D. Filev and H. Ying

8. Bonissone, P.: Fuzzy logic controllers: An industrial reality. In: Zurada, J., Marks, R.,
Robinson, C. (eds.) Computational Intelligence Imitating Life, pp. 316–327. IEEE Press, New
York, NY (1994)
9. Farinwata, S., Filev, D., Langari, R. (eds.): Fuzzy Control: Synthesis and Analysis. John Wiley
& Sons Ltd, London (2000)
10. Filev, D., Syed, F.: Applied control systems: blending fuzzy logic with conventional control
systems. Int. J. General Syst. 39, 395–414 (2010)
11. Eciolaza, L., Taniguchi, T., Sugeno, M., Filev, D., Wang, Y.: PB model-based FEL and its
application to driving pattern learning. In: 19th IFAC Congress, Cape Town, Aug 2014
12. Palm, R.: Robust-control by fuzzy sliding mode. Automatica 30(9), 1429–1437 (1994)
13. Palm, R., Driankov, D., Hellendoorn, H.: Model Based Fuzzy Control: Fuzzy Gain Schedulers
and Sliding Mode Fuzzy Controllers. Springer-Verlag, New York (1996)
14. Yager, R., Filev, D.: Essentials of Fuzzy Modeling and Control. John Wiley & Sons, New
York (1994)
15. Ying, H.: Practical design of nonlinear fuzzy controllers with stability analysis for regulating
processes with unknown mathematical models. Automatica 30, 1185–1195 (1994)
16. Ying, H.: Fuzzy Control and Modeling. IEEE Press, New York (2000)
17. Ying, H., Filev, D.: Analytical structure characterization and stability analysis for a general
class of mamdani fuzzy controllers. In: Zadeh, L.A., Abbasov, A.M., Yager, R.R.,
Shahbazova, S.N., Reformat, M.Z. (eds.) Recent Developments and New Directions in Soft
Computing, pp. 233–247. Springer Verlag (2014)
18. Du, X.Y., Zhang, N.Y., Ying, H.: Structure analysis and system design for a class of Mamdani
fuzzy controllers. Int. J. Gen Syst 37, 83–101 (2008)
19. Takagi, T., Sugeno, M.: Fuzzy identification of systems and its applications to modeling and
control. IEEE Trans. Syst. Man. Cybern. 15, 116–132 (1985)
20. Filev, D.: Fuzzy modeling of complex systems. Int. J. Approximate Reasoning 5, 281–290
(1991)
21. Filev, D.: Gain scheduling based control of a class of TSK systems. In: Farinwata, S., Filev,
D., Langari, R. (eds.) Fuzzy Control: Synthesis and Analysis, pp. 321–334. John Wiley &
Sons Ltd, London (2000)
22. Tanaka, K., Wang, H.O.: Fuzzy Control Systems Design and Analysis: A Linear Matrix
Inequality Approach. John Wiley & Son, New York (2001)
23. Bernstein, D.: From the editor—Theorem and look-up table. IEEE Control Syst. 24, 6 (2004)
24. Vogt, M., Muller, N., Isermann, R.: On-line adaptation of grid-based look-up tables using a
fast linear regression technique. Trans. ASME 126, 732–739 (2004)
25. Filev, D., Ying, H.: The look-up table controllers and a particular class of mamdani fuzzy
controllers are equivalent—implications to real-world applications. In: Proceedings of 2013
IFSA World Congress and NAFIPS Annual Meeting, Edmonton, Canada, 24–28, June 2013
26. de Boor, C.: A Practical Guide to Splines. Springer-Verlag (1978)
27. Ordonez, C., Omiecinski, E., Navathe, S., Ezquerra, N.: A clustering algorithm to discover low
and high density hyper-rectangles in subspaces of multidimensional data. Technical Report;
GIT-CC-99-20, Georgia Institute of Technology (1999)
28. Mizumoto, M.: Realization of PID controls by fuzzy control methods. In: Proceedings of IEEE
International Conference on Fuzzy Systems, San Diego, CA, 8–12 Mar 1992
29. Taniguchi, T., Sugeno, M.: Stabilization of nonlinear systems with piecewise bilinear models
derived from fuzzy if-then rules with singletons. In: Proceedings of the WCCI 2010 IEEE
World Congress on Computational Intelligence, pp. 2926–2931, Barcelona, Spain, 18−23 Jul
2010
30. Slotine, J.J., Li, W.P.: Applied Nonlinear Control. Prentice Hall (1991)
31. Wong, P.K., Tam, L.M., Li, K., Wong, H.C.: Automotive engine idle speed control
optimization using least squares support vector machine and genetic algorithm. Int. J. Intell.
Comput. Cybern. 1, 598–616 (2008)

[Link]@[Link]