ISSN 1064-2307, Journal of Computer and Systems Sciences International, 2006, Vol. 45, No. 5, pp. 744–771.

© Pleiades Publishing, Inc., 2006.

Original Russian Text © L.V. Litvintseva, S.V. Ul’yanov, S.S. Ul’yanov, 2006, published in Izvestiya Akademii Nauk. Teoriya i Sistemy Upravleniya, 2006, No. 5, pp. 69–97.

ARTIFICIAL
INTELLIGENCE

Design of Robust Knowledge Bases of Fuzzy Controllers

for Intelligent Control of Substantially Nonlinear Dynamic
Systems: II. A Soft Computing Optimizer and Robustness
of Intelligent Control Systems
L. V. Litvintsevab, S. V. Ul’yanova, and S. S. Ul’yanovb
aYamaha Motor Co., Ltd; Shizuoka, Iwata, Japan
bOOO MKG Kvant
Received January 19, 2006; in final form, March 13, 2006

Abstract—The structure of intelligent control system (ICS) is analyzed, and the interrelations with conven-
tional problems of the theory and practice of application of control systems are described. The analysis of the
results of simulation of typical structures of intelligent control systems has allowed us to establish the following
fact. The application of the technique of designing (presented in Part I), which is based on a fuzzy neural net-
work (FNN), does not guarantee in general that the required accuracy of approximation of the training signal
(TS) will be reached. As a result, under an essential change of external conditions, the sensitivity level of the
controlled plant (CP) increases, which, on the whole, leads to a decrease in the robustness of the intelligent con-
trol system, and, as a consequence, to a loss of reliability (accuracy) of achieving the control goal. To eliminate
the specified drawback of the neural network, a soft computing optimizer (SCO), which uses the technique of
soft computing and allows one to eliminate the drawback, is applied, which results in an increase in the robust-
ness level of the structure of the intelligent control system. The structure of the soft computing optimizer, which
contains as a particular case the required configuration of an optimal fuzzy neural network, is considered. The
main specific features of the functional operation of the soft computing optimizer and the stages of the process
of designing robust knowledge bases (KB) of fuzzy controllers (FC) are described. The methodology of joint
stochastic and fuzzy simulation of automatic control system based on the developed tool of the soft computing
optimizer is discussed in order to test the robustness and to estimate the limiting structural capabilities of intel-
ligent control systems. The efficiency of the control processes with application of the soft computing optimizer
is demonstrated by particular typical examples (benchmarks) of models of dynamic controlled plants under the
conditions of incomplete information about the parameters of the structure of the controlled plant and under the
presence of unpredicted (abnormal) control situations. Examples of industrial application of robust intelligent
control systems in actual control systems designed based on the soft computing optimizer are presented. Prac-
tical recommendations for improving the robustness level of intelligent control systems by using new types of
computations and simulation are given
DOI: 10.1134/S106423070605008X

INTRODUCTION plant [1]. This design methodology was based on the

theory of linguistic approximation and fuzzy inference
Engineering methods of control theory and the tech- (L. A. Zadeh and others) for developing robust knowl-
nique of designing automatic control systems were edge bases of intelligent fuzzy controllers [2, 3]. Within
formed in the past century. In particular, the founda- the framework of the specified methodology of design-
tions of stochastic and adaptive control of complex ing control laws based on physical methods (informa-
dynamic systems (in general, with variable structure) tion–thermodynamic and quantum–relativistic meth-
under conditions of information uncertainty were ods of describing controlled plants and control pro-
developed. cesses [4]), in the mid 1970s, the background of the
The next step in this direction was the development technique of designing intelligent automatic control
of principles of simulation and design of fuzzy systems was developed [5–8].
automatic control systems under uncertainty conditions Remark 1. One of the most important elements of
that take into account the individual specific features of science intensive information technology of designing
the behavior of chosen trajectories of the controlled intelligent control systems is the development of the

744
 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 745

methodology and the corresponding software–hard- chosen trajectories) of stochastic interactions with
ware support of the processes of testing and evaluating given characteristics (see the Appendix). In the course
the robustness level of the designed structure of the of optimization of the parameters of automatic control
intelligent control system (as a measure of sensitivity to systems, this approach of stochastic simulation is used
various external and internal random perturbations act- together with the method of fuzzy simulation in order
ing both on the controlled plant and in the measurement to achieve the required control quality irrespective of
channels or control loops). The topicality of the solu- the particular implementation of the perturbing sto-
tion to this problem is dictated by practical control tasks chastic action. We consider the methodology of joint
addressed by many researchers [2, 3, 9–12]. stochastic and fuzzy simulation of automatic control
An increase in the complexity of the structures of systems based on the developed tools of the soft com-
controlled plants and difficulties in predicting unex- puting optimizer with the aim to test the robustness and
pected (abnormal) control situations only stresses the estimate the limiting structural capabilities of intelli-
topicality of this problem and draws attention to it. gent control systems. The efficiency of control pro-
Such problems are referred to as the so-called problem cesses with application of the soft computing optimizer
of “System of Systems Engineering” that studies in a is demonstrated by particular typical examples of mod-
general form complex structures of automatic control els of dynamic controlled plants under the conditions of
systems with different levels and scales of integration uncertainty of information about the parameters of the
and/or priority data exchange between subsystems in controlled plant and under the presence of unpredicted
order to establish global (necessary and sufficient) con- (abnormal) control situations.
ditions for reliable autonomous operation of the con- Thus, in connection with the fact that within the
trolled plant in the external environment. framework of the classical approach to designing auto-
Remark 2. In a number of cases, in the experience matic control systems it is not possible to improve the
of designing automatic control systems, linearized control quality and the robustness level of the control
models of controlled plants are applied. However, laws obtained, the problem of development of methods
under this approach, the adequacy of the relation of mathematical simulation of algorithms for intelligent
between the physical parameters of the controlled plant control of nonlinear dynamic systems based on soft
and the parameters of its linearized model is frequently computing and software tools for their support still
lost. In this paper, in the optimization of the structure remains topical. It is these problems to which this paper
and the parameters of an intelligent control system, we is dedicated.
consider nonlinear models of controlled objects. Note Since the main point in the design of intelligent con-
that the effect of nonlinearities on the dynamics and on trol systems is the stage of forming the corresponding
the controllability of the controlled object in the control knowledge base [1, 3], the design of a knowledge base
laws is taken into account by the gains of a fuzzy PID with a required robustness level, under unpredicted
controller. Thus, by methods of soft computing, the control situations, allows one to establish in a general
effect of nonlinearities in the controlled plant is com- form the correspondence between the conditions of
pensated by varying dynamically the proportionality functioning of the controlled plant and the required
coefficients in the classical PID controller with global robustness level of the intelligent control system.
negative feedback (GNF). In this paper, we focus the main attention on the
In practice, controlled plants are under conditions of description of particular results of designing knowl-
uncertainty connected with the effect of both external edge bases and simulating intelligent control systems
and internal random factors. The ability of automatic with essentially nonlinear controlled plants with a ran-
control systems to react adequately to one or another domly variable structure and stimuli (control goals). In
change in the parameters of the external environment this case, the aim of this work is to determine the
that is not given in advance in designing the automatic robustness levels of control processes that ensure the
control system characterizes the level of learning, adap- required reliability and accuracy indices under the con-
tation, and robustness of processes and the ability to ditions of uncertainty of the information employed in
learn and adapt itself. Frequently, the methods of the decision-making. First of all, we consider the evolution
theory of robust control are not able to solve general of typical structures of intelligent control systems, their
control problems under the presence of uncertainty specific features, advantages, and disadvantages from
described in the form of a certain stochastic process with the point of view of design and application of intelli-
definite (in general, unknown) statistical characteristics gent control systems.
(probability distribution functions). A number of
approaches to the solution of this problem based on iter-
ative randomized algorithms were developed in [10, 11]. 1. SPECIFIC FEATURES OF DESIGNING
THE STRUCTURES OF ROBUST INTELLIGENT
In this paper, the evaluation of sensitivity and the CONTROL SYSTEMS
improvement of the robustness level is reached, in par-
ticular, by using algorithms of nonlinear forming filters One of the main tasks of designing an intelligent
for reproducing realizations (informatively represented control system consists in providing that the developed

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

746 LITVINTSEVA et al.

(chosen) structure possesses the required level of con- to establish in an analytic form the correspondence
trol quality and robustness (supports the required indices between the required level of stability, controllability,
of reliability and accuracy of control under the condi- and robustness of the control [1–3, 8, 12]. This allows
tions of information uncertainty). Note that one of the one to determine the required intelligence level of the
most important and hard-to-solve problems of designing automatic control system depending on the complexity
intelligent control systems is the design of robust knowl- of the particular control problem.
edge bases that allow the intelligent control system to Example 1. Let us consider in short the main phys-
operate under the conditions of information uncertainty. ical principles of control processes that allow one to
The core of technique for designing robust knowledge establish the interrelation between the qualitative
bases of fuzzy controllers is generated by new types of dynamic characteristics of the controlled plant and the
computing and simulation processes [3, 7, 9]. actuator of the automatic control system: stability, con-
Recently, the application of the structures of intelli- trollability, and robustness of control. For this purpose,
gent control systems based on new types of computa- we employ the informational and thermodynamic
tions (such as soft, quantum, etc., computing) has approaches that join by a homogeneous condition the cri-
drawn the ever-increasing attention of researchers [1–9, teria of dynamic stability (the Lyapunov function), con-
11–13]. Numerous investigations conducted have trollability, and robustness [2]. Consider a dynamic con-
shown that they possess the following points of favor: trolled plant given (in a general form) by the equation
retain the main advantages of conventional auto- dq
matic control systems (such as stability, controllability, ------ = ϕ ( q, S ( t ), tu ) (1.1)
observability, etc.); dt
have an optimal (from the point of view of a given where q is the vector of generalized coordinates
control objective functional) knowledge base, as well describing the dynamics of the controlled plant; S(t) is
as a possibility of correcting it and adapting it to the the generalized entropy of dynamic system (1.1); u is
changing control situation; the control force (the output of the actuator of the auto-
guarantee the attainability of the required control matic control system); and t is the time. The necessary
quality based on the designed knowledge base; and sufficient conditions of asymptotic stability of
dynamic system (1.1) are determined by the physical
are open systems, i.e., they allow one to introduce an constraints on the form of the Lyapunov function,
additional objective functional for control and con- which possesses two important properties represented
straints on the characteristics of the control process. by the following conditions:
One of the main problems of modern control theory (1) This is a strictly positive function of generalized
is to develop and design automatic control systems that coordinates, i.e., V > 0.
meet the three main requirements: stability, controlla-
bility, and robustness [2, 3, 9–11]. The listed quality cri- (2) The complete derivative in time of the Lyapunov
dV
teria ensure the required accuracy of control and reliabil- function is a nonpositive function, ------- ≤ 0.
ity of operation of the controlled plant under the condi- dt
tions of incomplete information about the external By conditions (1) and (2), as the generalized
perturbations and under noise in the measurement and Lyapunov function, we take the function [12]
control channels, uncertainty in either the structure or n
parameters of the controlled plant, or under limited pos-
∑q
1 2 1 2
sibility of a formalized description of the control goal. V = --- i + --- S , (1.2)
2 2
This problem is solved in three stages. i=1

(1) The characteristics of stability of the controlled where S = Sp – Sc is the production of entropy in the
plant are determined for fixed conditions of its opera- open system “CP + controller”; Sp = Ψ(q, q̇ , t) is the
tion in the external environment. production of entropy in the controlled plant; and Sc =
(2) A control law is formed that provides the stabil- ϒ( ė , t) is the production of entropy in the controller
ity of operation of the controlled plant for a given accu- (actuator of the automatic control system). It is possible
racy of control (according to a given criteria of the opti- to introduce the entropy characteristics in Eqs. (1.1)
mal control). and (1.2) because of the scalar property of entropy as a
(3) The sensitivity and robustness of the dynamic function of time, S(t) [2].
behavior of the controlled plant are tested for various Remark 3. It is worth noting that the presence of
classes of random perturbations and noise. entropy production in (1.1) as a parameter reflects the
These design stages are considered by modern control dynamics of the behavior of the controlled plant and
theory as relatively independent. The main problem of results in a new class of substantially nonlinear
designing automatic control systems is to determine an dynamic automatic control systems. The choice of the
optimal interaction between these three quality indices. minimum entropy production both in the controlled
For robust structures of automatic control systems, a plant and in the fuzzy PID controller as a fitness func-
physical control principle can be proven that allows one tion in the genetic algorithm allows one to obtain feasi-

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 747

ble robust control laws for the gains in the fuzzy PID edge bases of robust intelligent control systems (with
controller [8]. The entropy production of a dynamic different levels of intelligence [2, 3]) with the use of
system is characterized uniquely by the parameters of soft computing. In concluding this section, we formu-
the nonlinear dynamic automatic control system, which late the following conclusions.
results in determination of an optimal selective trajec- 1. The introduced physical law of intelligent control
tory from the set of possible trajectories in optimization (1.3) provides a background of design of robust knowl-
problems [2, 3, 12]. edge bases of intelligent control systems (with different
The first condition is fulfilled automatically. levels of intelligence) based on soft computing.
dV 2. The technique of soft computing gives the oppor-
Assume that the second condition ------- ≤ 0 holds. In this
dt tunity to develop a universal approximator in the form
case, the complete derivative of the Lyapunov functi- of a fuzzy automatic control system, which elicits
on (1.2) has the form information from the data of simulation of the dynamic
n
behavior of the controlled plant and the actuator of the
automatic control system.
∑ ∑ q̇ q + SṠ
dV 1 1
------- = --- 2q̇ i q i + --- 2SṠ = i i 3. The application of soft computing guarantees the
dt 2 2
i=1 purposeful design of the corresponding robustness level
n by an optimal design of the total number of production
= ∑ q ϕ ( q, S, t, u ) + ( S
i p – S c ) ( Ṡ p – Ṡ c ). rules and types of membership functions in the knowl-
edge base [3, 18].
i=1

Thus, taking into account (1.1) and the notation intro- Figures 1 and 2 present typical criteria for control
duced above, we have quality, their interrelations with different types of com-
putations and simulation types, as well as the hierarchy
dV of levels of control quality depending on the required
-------
dt level of intelligence of the automatic control system.
⎧
⎨
⎩

stability (1.3) Figure 3 presents the main components and their

n
interrelations in the information design technology
= ∑ q i ϕ ( q, ( Ψ – ϒ ), t, u ) + ( Ψ – ϒ ) ( Ψ̇ – ϒ̇ ) ≤ 0 . (IDT) based on new types of computing (soft and quan-
i=1 tum). The key point of this information design technol-
⎧
⎪
⎪
⎨
⎪
⎪
⎩
⎧
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎩

controllability robustness ogy is the use of the method of eliciting objective

Relation (1.3) is a generalization of the results of [1, knowledge about the control process irrespective of the
8, 12] and relates the stability, controllability, and subjective experience of experts and the design of
robustness properties. objective knowledge bases of a fuzzy controller [3, 18],
Remark 4. The presence of entropy production in which is principal component of a robust intelligent
the motion equations of the controlled plant is the prin- control system. The output result of application of this
cipal distinction of (1.3) from the corresponding for- information design technology is a robust knowledge
mula of [1–3, 8, 12]. The practical application of con- base of the fuzzy controller that allows the intelligent
trol law (1.3) to conventional problems such as the esti- control system to operate under various types of infor-
mation of accuracy (roughness) of the linearization of mation uncertainty.
models of controlled plants, observability of the param-
eters of control processes, etc., was considered in [2, 3, 2. EVOLUTION OF OPTIMIZATION PROCESSES
12]. We note that the term ∑
q i q̇ i characterizes an AND STRUCTURAL ANALYSIS OF
INTELLIGENT CONTROL SYSTEMS
i
additional opportunity to work with a physical model of Optimal control problems for substantially nonlin-
the controlled plant without a mathematical model, ear and globally unstable controlled plants were studied
using directly the measurement of the indices of under the presence of stochastic perturbations of vari-
dynamic behavior of the controlled plant. In this case, ous types (a set of typical essentially nonlinear oscilla-
we have a generalization of the black-box model of the tors were considered as controlled plants). With the
controlled plant. Other approaches of entropy analysis help of the developed system of stochastic simulation,
of dynamic systems were described in [14–17]. the limiting capabilities of classical automatic control
Equation (1.3) joins in an analytic form different systems based on the use of the actuator in the form of
measures of control quality such as stability, controlla- a conventional PID controller were determined
bility, and robustness supporting the required reliability (Table 1, position 1). As a result of the conducted inves-
and accuracy. Consequently, the interrelation between tigation, the following fact was established: classical
the Lyapunov stability and robustness described by automatic control systems based on a PID controller
Eq. (1.3) is the main physical law for designing auto- with constant gains frequently do not cope with the
matic control systems. This law provides the back- control problem in the case of globally unstable and
ground for an applied technique of designing knowl- essentially nonlinear controlled plants subjected to

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

748 LITVINTSEVA et al.

Levels Control quality

0 Reliability

1 Stability Robustness Controllability

Training Adaptation
2
Soft
computing
System
engineering
Engineering
design
problem
SCO
KB ëë
FC.
3 Self-organization
PID CP
Quantum soft
computing m(t)
m(t)
m(t)

Fig. 1. The interaction between the types and hierarchical levels of control quality criteria.

non-Gaussian (e.g., Rayleigh) statistical noise, as well trol strategies formulated by a human expert (in com-
as under the presence of time delays in the measure- plex unpredicted situations of control and decision
ment and control channels. making).
This fact can be explained by the following circum- In this paper, the control quality indices are consid-
stances. Classical automatic control systems are based ered in this case as components of vector fitness func-
on the global-negative-feedback principle and the tion of the corresponding genetic algorithm in the soft
method of compensation of the control error. However, computing optimizer.
in complex controlled plants (mentioned above), it is
not sufficient to take into account the result of optimi-
zation according to only one quality control index, such 2.1. Generations and Evolutions of the Typical
as the minimum of control error. Therefore, in the prac- Structures of Intelligent Control Systems
tice of designing intelligent control systems for com- Table 1 presents the generalized evolution of the
plex controlled plants, the following problem arises: development and formation of the structures of intelli-
How will we introduce in the control system other addi- gent control systems, their specific features, and their
tional control quality indices? advantages and disadvantages, as well as the levels of
For example, among these criteria we have the crite- control quality belonging to these structures.
rion of minimum entropy production in the automatic At the first stage, for simulating human–machine
control system itself mentioned above (taking into control strategies, so-called soft computing was used.
account the heat loss and loss of useful work in the Soft computing is based on set theory and fuzzy infer-
object and control system), as well as other more com- ence [1–3]. For example, the structure of position 1 of
plex vector control quality indices. Table 1 is transformed to an expert control system (ES)
It is practically impossible to solve this problem by growing step by step the constituent blocks in the
with the help of design of a standard PID controller classical version of the automatic control system (in
with constant gains. The limited opportunities of real- this case, by introducing a block of fuzzy inference)
ization of complex quality control indices in classical (Table 1, position 2) [2] and is an example of the first
automatic control systems, especially in the control generation of intelligent control systems.
cases specified above, provide a stimulus for develop- Thus, the first generation of intelligent control sys-
ing intelligent control systems based on the use of con- tems was considered as expert systems with knowledge

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 749

Tools ?
Control quality
Accuracy Controllability Stability
1. Controllability
Reference + ε Conventional Controlled Accuracy
2. Stability
signal controller plant
– 3. Learning
Robustness
4. Adaptation

Types of control Technology of design

Requirements
& quality [based on new types of computing]
Constraints
Stability Integer calculus
Controllability Noninteger calculus
CP Stochastic calculus
Robustness
Adaptation Soft computing
Learning Quantum computing
Self-organization Quantum soft computing

Laboratory Experimental Industrial Industrial

testing specimen application specimen

Reliability
Accuracy Technology of design
and implementation of the ICS Robustness
Robustness Accuracy

Fig. 2. The interrelation between the control quality criteria, types of intelligent computing, and simulation in designing robust KBs
of the FC.

representations different in depth. The main role in control systems of the second generation were devel-
these expert systems was played by the quality of the oped with a deeper knowledge representation, using the
knowledge base, which depends on the experience and technique of so-called soft computing, which joins in a
subjective knowledge of the human expert. However, in single chain genetic algorithms, fuzzy neural networks,
the case when the control of globally unstable and sub- and fuzzy controllers. This has allowed one to eliminate
stantially nonlinear controlled plants is subjected to the subjective expert opinion at the stage of forming the
complex stochastic noise, it is difficult even for an structure and parameters of the knowledge base of the
experienced human operator to choose an optimal fuzzy controller. In turn, the development of intelligent
(from the point of view of control quality) knowledge control systems based on soft computing has generated
base of the fuzzy controller. This problem is a bottle- several approaches to forming the structures of knowl-
neck of all first (and the following modifications) intel- edge bases. It was planned initially to form a certain
ligent control systems [2, 3]. Therefore, the use of rather rough training signal (TS) from the genetic algo-
expert systems as a tool for knowledge acquisition and rithm and to acquire the knowledge base by approxi-
formation of knowledge bases (as the background for mating the obtained training signal on fuzzy neural net-
designing intelligent control systems) has not resulted works. The structure of intelligent systems of the sec-
in the expected essential success (although there are ond generation is formed by introducing a block
many example of industrial application [5, 19]) because containing genetic algorithms and fuzzy neural net-
of the complexity of control plants and the subjective works in the structure of intelligent control systems of
nature of the information provided by an expert. the first generation (Table 1, position 3).
From the point of view of design technology, the Then, the second generation of intelligent control
main problem of application of intelligent control sys- systems began to use a new form of feedback, called a
tems of the first generation was their weak adaptivity to global intelligent feedback (GIF) [1, 3, 8, 12, 18] repre-
changes in the parameters of controlled plants (e.g., sented in Table 1 (position 4). This gives the opportu-
caused by the fact that the structure of the controlled nity to acquire objective knowledge directly from the
plant is getting older or by sharp change of the external dynamic behavior of the controlled plant and the actu-
environment), as well as the low robustness of control ator of the automatic control system. The GIF loop con-
laws obtained. For solving such problems, intelligent tains a genetic algorithm for obtaining information

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

750 LITVINTSEVA et al.

Basic element for designing ICS

with different levels of intelligence
Description of the controlled plant.
Various stochastic
perturbations Levels of control characteristics
Basic level Intelligence Robustnes Control quality

Control system based on Problem

the classical theory of simple control ?
systems Technology of designing ICSs

Robust KB Structures of ICSs

New technology
Works Does not work: Technique of soft computing
Rayleigh noise, FC based on certain original
principles of the genetic algorithm
nonlinear CP Step 1
Technology problems Current results
Works:
Proposals for the future Existing difficulties Possibilities of simplification,
improvement of robustness
Soft computing optimize Difficulties in finding
an optimal solution Application in practice
Technique of quantum computing Guarantee of robustness
of complex stochastic
perturbations Step 3
Step 2
Solved problems Quantum soft computing

Fig. 3. The process of development and creation of information technology for designing an integrated intelligent control system.

about the optimal control signal (based on the dynamic metric and external random perturbations of various
and thermodynamic behavior of the controlled plant probabilistic natures (with different density functions
and the PID controller) and a fuzzy neural network of probability distribution). Simulating the behavior of
approximating a given optimal control signal with the the controlled plant on the set of chosen typical sub-
help of a given structure of the neural network. This stantially nonlinear oscillators, the limiting capabilities
approach was considered in [1]. of the intelligent control system were conducted. As a
The main block in the structure of the intelligent result, it was established that in the case of instable and
control system (Table 1, position 5) is a system of sim- substantially nonlinear controlled plants, as well as
ulating the optimal control signal (SSOCS) with the under the action of Rayleigh stochastic noise on them,
help of a genetic algorithm and a control quality index it is provided that a certain robustness level is reached
represented in the form of one of the components of the for particular classes of random perturbations. How-
fitness vector function of the genetic algorithm. The ever, we cannot manage to design a robust knowledge
output of the global intelligent feedback is a training base using only stage 1 of the design of the knowledge
signal of the optimal control in the form of the follow- base of the intelligent control system (Fig. 3) [3, 18].
ing input/output data: {E(ti), K(ti)}, i = 1,…n, where
The analysis of results of simulation of the auto-
∫
E(ti) = {e(ti), ė (ti), e ( t i ) dt i } is the vector whose com- matic control system based on the first stage of the
developed technique of designing knowledge bases has
ponents are the control error, its derivative, and the inte-
gral of the error, respectively; K(ti) = {kp(ti), kd(ti), ki(ti)} shown that the main drawback of this stage is the non-
optimal choice of the structure of the fuzzy neural net-
are optimal (from the point of view of a given fitness
work approximating the training signal. As a rule, in the
function of the genetic algorithm) parameters of the
system for designing knowledge bases of intelligent
PID controller; and ti is the time instant.
control systems of this type, the design of the corre-
Using the training signal and the mechanism of sponding structure of the fuzzy neural network is the
supervisor training of the fuzzy neural network based duty of an experienced human operator. In this case, the
on the error-back-propagation method, we can design training signal is divided into input and output compo-
the knowledge base of the fuzzy controller represented nents, each of which, in turn, consists of one or more
by a given fuzzy neural network. This stage is denoted signals. In the general form, each of the investigated
as step 1 in Fig. 3 and was described in detail in [1]. signals is described by a selective (representative) tra-
The main disadvantage of intelligent control sys- jectory of a certain stochastic process. Note that it is
tems with GIF consists in the possibility of reaching the kept in mind that, at each time instant, there is a depen-
required robustness level on a given class both of para- dence between the input and output signals. For exam-

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 Table 1. The evolution of the process of designing the structures of knowledge-based ICS and interrelation with classical control theory

Structures of control systems and evolution of development Control quality levels Advantages Limit capabilities

Position 1 Base element of the ACS Stability Simplicity of simulation and Under a change of external conditions
(Y) + ε + (X)
Controllability physical implementation (min or the complexity of the CP, does not
PID U* Controlled
controller plant
Control precision complexity) provide sufficient conditions
Reference – u + Simplicity in HW implemen- Stability and controllability (the ab-
signal tation sence of the required robustness level).
Under minimum complexity Rigid (nonadaptive) structure
m(t) of CP provides minimum sta- Does not possesses an ability to learn
bility and controllability and be self-organized
Measurement
system Guarantees the necessary con-
ditions of control quality, but
not sufficient
Knowledge-based intelligent control systems
Position 2 Base element of the ICS Stability Simplicity and efficiency Choice of membership function
Controllability of control of complex CPs depends of the qualification level
Control precision Can be easily designed and of the expert
K Expert system
FC Partial adaptation implemented Has limited capabilities in practical
High reliability in exploita- implementations because of constraints
tion on the description of fuzzy relations
(Y) + e PID + U * Controlled (X)
Is used in designing robust “input–output”
controller u plant KBs Problem of complexity of designing
Reference – + a KB of the FC since there is no
signal
algorithm for constructive generation
m(t) of KBs

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL

Requires large amount of time for
Measurement system constructing logical rules of fuzzy
inference

Vol. 45
Position 3 Block of knowledge extraction and generation of KBs Stability Allows one to correct partially Does not possess sufficient robustness
Controllability the drawbacks of the basic and stability. Uses information about
Control precision level of the ICS and to achieve control error and does not use informa-
GA FNN FC
Adaptation the required adaptation level tion about dynamic behavior of the FC

No. 5
under fixed conditions of and CP
ä functioning of the CP
DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS

2006
(Y) + ε PID + U* Controlled (X)
Reference controller plant
– u +
signal

m(t)

Measurement system
751
 Table 1. (Contd.)
752

Structures of control systems and evolution of development Control quality levels Advantages Limit capabilities

Position 4 Intelligent feedback Stability Intelligent feedback allows FNN guarantees the required accuracy
SSOCS Controllability one to acquire objective of approximation of the training signal
FNN GA Control precision knowledge and to construct from the output of the GA
ä Adaptation the KB of FC Under a sharp change of the external
FC Learning Allows one to encompass a conditions, improves the sensitivity of
Partial self-organization new level of quality control the CP (reduces robustness)
(Y) + ε + (X) (partial self-organization)
PID U * Controlled
Reference – controller u + plant
signal
m(t)

Measurement system
LITVINTSEVA et al.

Position 5 Soft computing optimizer Stability Retains the advantages of the Because of limit capabilities of the
SSOCS Controllability previous levels of the ICS and search algorithm, as a GA, can guaran-
SCO GA Control precision provides a background for the tee the control quality only on a fixed
Adaptation technology of designing ICSs. space for searching solutions
ä Learning Allows one to optimize the Process of designing KBs is performed
FC
Self-organization of a KB by the method of combinatorial search
given level Guarantees the attainability of on the decision tree
(Y) + ε PID + U * Controlled (X) the required control quality

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL

Reference – controller u plant based on the designed KB
+
signal Is an open system, which al-
m(t) lows one to introduce addi-
tional control objective func-
Measurement system

Vol. 45
tionals
Allows one to design DBs for
the subsequent stage of de-

No. 5
signing based on quantum
computing

2006
 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 753

ple, when the control signal is approximated, the con- the control loop and in the measurement system with-
trol error and its derivative (control error rate) may be out loss of accuracy and control quality [22].
the input components and either the required value of The robustness of intelligent control systems
the control action or certain adjusted parameters of the obtained based on this approach requires minimum
automatic control system (e.g., the gains of the PID source information both about the behavior of the
controller) may be the output components. The task of controlled plant and about the external perturbations
the expert in determining the structure of the fuzzy neu- [20, 21]. The system for simulation and design of the
ral network is reduced to the choice of the model of structures of intelligent control systems was devel-
fuzzy inference and, mainly, to a linguistic description oped based on the soft computing optimizer of the
of the given training signal. A linguistic variable that “GA−OKB–Fuzzy PID controller” type (Table 1, posi-
describes the signal by the term-set corresponding to tion 5).
this linguistic variable corresponds to each component
of the training signal. The cardinality of the term-set Analysis of the structures of the existing automatic
and the characteristics of its elements (the class and control systems has allowed us to choose as a basic
parameters of the membership function) are unknown. automatic system the conventional automatic control
The “completeness” of the linguistic description of the system in the form of a PID controller. This minimum
signal can be given at the level of the interrelation of the complex structure joins the maximum number of con-
term-sets belonging to linguistic variables. This prob- trol quality indices, i.e., stability, control accuracy, and
lem is also solved by a human in systems for designing controllability, guaranteeing a certain (minimum)
intelligent control problems based on conventional soft robustness level. Using the physical control law, which
computing (the second generation of intelligent control relates these criteria, it is possible to design intelligent
systems). However, as was stated above, under complex control systems that meet the requirements of control
control situations, it is difficult even for an experienced quality with a sufficient robust level.
expert to solve this problem manually (i.e., to choose an Thus, by introduction and interaction of a global
optimal structure of the fuzzy neural network for a negative feedback and a global intelligent feedback, the
given training signal). principle of saving the lower control level is imple-
Another important problem is to determine the mented in accordance with the hierarchy (priority) of
required relation between the accuracy of description levels of control quality (Fig. 1). This detects a bottle-
(approximation) of the training signal and the required neck in the structure of the intelligent control system,
robustness level of the whole structure of the fuzzy neu- which is the process of forming and designing the
ral network. Both specified problems are solved at the knowledge base of the fuzzy controller. To solve this
second stage of the technology of designing the knowl- problem, a technique for designing the knowledge base
edge base for an intelligent control system by the soft- (step 2) was developed for the unified substantiated
ware facilities of the tool called a soft computing opti- structure of the intelligent control system (Fig. 3) [3,
mizer (see Section 3) [3, 18, 20, 21]. 20, 21].
As follows from Table 1 (position 5), the advantages
of this structural level of the intelligent control system
2.2. Structural Analysis of Intelligent Control Systems consist in the fact that the global intelligent feedback
with the Help of a Soft Computing Optimizer allows one to design the knowledge base of the fuzzy
With the help of random search and natural selec- controller based on objective knowledge acquisition
tion (based on the structure of the genetic algorithm from analysis of the dynamic behavior of the controlled
developed by the authors), different variants of robust plant and the fuzzy controller itself. Note that because
knowledge bases were simulated. The formed robust of using the corresponding fitness function of the
knowledge bases allow one to control complex con- genetic algorithm (like entropy production rate as a
trolled plants in conditions of uncertainty of informa- physical optimality criterion, etc.), there arises an
tion about external perturbations acting on the con- opportunity to optimize the structure of the knowledge
trolled plant and the changes of reference signals (con- base itself. The global intelligent feedback helps to
trol goals). The robustness of control laws is reached by encompass new levels of control quality and elements
introducing vector fitness functions of the genetic algo- of self-organization simultaneously. The introduction
rithm, which contain as a component the physical prin- of the level of quality control such that learning allows
ciple of minimum production of the generalized us to improve the robustness of control and its stability
entropy both in the controlled plant and in the intelli- and, together with adaptation, to reduce the require-
gent controller [3, 8]. This approach allows one to ful- ments for the amount of source information about the
fill the principle of designing an optimal intelligent external condition of operation of the controlled plant.
control system with a maximum level of reliability and In addition, as a result of application of learning pro-
controllability for a complex controlled plant under the cesses, it is possible to reduce the requirements for
conditions of uncertainty of the source information; required energy expenditure (useful work) both in the
reduce to a required minimum the number of sensors controlled plant and in the structure of the conventional
for collection and transmission of information both in PID controller.

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

754 LITVINTSEVA et al.

Table 2. The types and the role of the fitness function of the GA in the SCO

Type of GA Criteria Fitness function (FF) Role of the FF

GA1: Maximum joint j j j j j Elimination

H X i = – p X i log ( p X i ) = – p ( x i x i = µ X i ) log [ p ( x i x i = µ X i ) ]
Optimization information of redundancy
of linguistic entropy and N of the TS
1
= – ---- ∑ µ X i ( x i ( t ) ) log [ µ X i ( x i ( t ) ) ]
variables minimum j j
Choice of an
information N optimal cardinality
t=1
about signals of term-sets
separately N of linguistic
= H⎛x j ⎞ = – ---
( j, l ) 1
∑
j l
H Xi - [ µ ( x ( t ) ) µ ( x ( t ) ) ], variables of the
Xk ⎝ x = µ j , x = µl ⎠ N X i i * X k k
components
i Xi k Xk t=1
of the TS
j l
log [ µ X i ( x i ( t ) * µ X k ( x k ( t ) ) ) ],
where * is the chosen operation of fuzzy AND
GA2: Minimum
∑E Choice of optimal
p
E = ,
Optimization approximation parameters of the
p
of rule base error right sides of rules
p p p p p 2
where E = 1/2 ( F ( x 1 , x 2 , …, x n ) – d )

GA3: Minimum Fine adjustment

∑E
p
Adjustment of approximation E = of the parameters
the KB error or maxi- p of membership
mum joint infor- j functions
mation entropy H Xi

The introduction of global intelligent feedback imentally, i.e., directly from the measurement of the
allows one to extract valuable information from the parameters of the physical model of the controlled
open system “CP + conventional controller.” To plant. Figure 4 also presents the successive implemen-
increase the robustness level, it is necessary to develop tation of the stages of designing the soft computing
the corresponding tools that support the formation of optimizer. Let us specify the steps of the optimization
the knowledge base of the fuzzy controller. Consider
the structure of the soft computing optimizer for form- algorithm.
ing and designing robust structures of intelligent con- Step 1. Choice of the model of fuzzy inference. The
trol systems. user specifies the particular type of model of fuzzy
inference (Sugeno, Mamdani, etc.) and the number of
3. THE STRUCTURE OF THE SOFT COMPUTING input and output variables.
OPTIMIZER Step 2. Creation of linguistic variables. With the
The soft computing optimizer is a new, efficient help of genetic GA1, an optimal number of membership
software tool for designing knowledge bases of robust functions is determined for each input linguistic vari-
intelligent control systems based on soft computing able, and an optimal form for the representation of its
with the use of new optimization criteria (in the form of membership functions (triangular, Gaussian, etc.) is
new fitness functions of genetic algorithms). As these chosen.
criteria, we take the thermodynamic and information–
entropy criteria [2] represented in Table 2. Step 3. Design of the rule base. At this stage, a spe-
The structure of the soft computing optimizer [18, cial algorithm for selection of the most robust rules is
21] for designing robust intelligent control systems is used in accordance with the following two criteria:
presented in Fig. 4. The soft computing optimizer con-
sists of interrelated genetic algorithms (GA1, GA2, (1) “total” criterion: choose only the rules that sat-
GA3), which optimize particular components of the isfy the following condition:
knowledge base.
R total_ fs ≥ TL,
l
The input of the soft computing optimizer is a train-
ing signal, which can be obtained either at the stage of
stochastic simulation of the behavior of the controlled where TL (threshold level) is a given (manually or cho-
plant (with the use of its mathematical model) or exper- sen automatically) level of rule activation, and

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 755

Input: training signal

SCO
Step 1 Step 2
Choice of a model of fuzzy
GA1 creation of linguistic
inference (Sugeno, Mamdani,
variables
etc.)

Step 3 Step 4
Creation of the predicate part
of the rule base: GA2 optimization of the rule
“total” criterion, base
“maximum” criterion

Adjustment of the KB (GA3: optimization

of linguistic variables Step 5

Step 5.1 Step 5.2 Step 5.3

According to the criterion According to the criterion Fine adjustment
of approximation-error of parameters of the KB
of maximum of information
minimum entropy by gradient methods

Output: a KB of the FC

Fig. 4. The algorithm of interaction of operations in the soft computing optimizer based on soft computing.

N help of the next step, this solution can be improved

∑ R fs ( t k ), and
l l locally.
R total_ fs =
k=1 Step 5. Adjustment of the base of rules. With the
help of GA3, the left and right sides of the rules of the
R fs ( t k ) = Π [ µ j1 ( ( x 1 ( t k ) ), µ j2 x 1 ( t k ) ), …,µ jn ( x n ) ( t k ) ) ],
l l l l
knowledge base are optimized; i.e., optimal parameters
of the membership functions of the input/output vari-
where tk are time instants, k = 1, …, N, and N is equal to ables are chosen (from the viewpoint of a given fitness
the number of points in the control signal; µ jk (xk), k = 1,
l function of the genetic algorithm). In this optimization
…, n are membership functions of input variables, l is process, three different fitness functions chosen by the
the index of the rule in the knowledge base; and symbol user (steps 5.1 and 5.2 in Fig. 4) are used. In addition,
“Π” means the operation of fuzzy conjunction (in par- there is also the opportunity to adjust the knowledge
ticular, it may be interpreted as a product); base with the help of conventional error-back-propaga-
tion method (step 5.3 in Fig. 4).
(2) “maximum” criterion: choose only the rules that Verification (testing) of the designed knowledge
satisfy the condition base. Constructed at stages 4, 5.1, 5.2, and 5.3 knowl-
edge bases of the intelligent control system are tested
max R fs ( t ) ≥ TL.
l
from the viewpoint of robustness and control quality.
t
For further use, the functionally best knowledge base is
Step 4. Optimization of base rules. With the help of chosen, which is tested in the functioning mode in real
GA2, the right sides of rules of the knowledge base time.
defined at step 3 are optimized. At this stage, a solution Examples of simulation of knowledge bases on the
that is close to the global optimum is found (minimum basis of efficient application of the soft computing opti-
error of approximation of the training signal). With the mizer are considered in Section 5.

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

756 LITVINTSEVA et al.

Fig. 5. The main menu of the soft computing optimizer.

4. SOFTWARE IMPLEMENTATION OF THE SOFT accurate approximation of the training signal by the
COMPUTING OPTIMIZER obtained rules (Refine KB). The error-back-propaga-
The soft computing optimizer was implemented as a tion algorithm is also included (Back propagation),
software system [3, 18, 21]. As a programming lan- which guarantees a given accuracy of the approxima-
guage, C++ (Microsoft Visual Studio.net) was chosen. tion of the training signal of the designed knowledge
The algorithmic part devoted to the implementation of base.
the main stages of optimization algorithms was imple- In the central section of the main menu of the opti-
mented as a platform-independent tool. The graphical mizer, the basic information about the designed fuzzy
interface presented in Fig. 5 was developed for operat- system is located, such as the type, address of the main
ing systems of the Win32 family and was tested on per- file of the knowledge base, the number of input and out-
sonal computers with different versions of the Windows put variables, as well as generic information about the
operating system. training signal. Here, you can also find the editor of lin-
The main menu of the optimizer was divided into guistic variables and the editor of rules.
several sections (Fig. 5) devoted to execution of the Figure 6 presents the editor of linguistic variables.
main functions and visualization of the results of algo- The membership functions of fuzzy variables can be
rithm operation. edited both manually, by dragging the corresponding
In the left section of the main menu, a group of but- values, and by manual input of parameters.
tons is located. These buttons run different optimizing Figure 7 presents the editor of the base of fuzzy
components, such as rules. The fuzzy rules are structurally represented in the
creation of linguistic variables (Create variables) form of a fuzzy neural network. The number of neurons
with the help of GA1; of the first layer corresponds to the number of input sig-
nals, while the number of neurons of the second layer
algorithm of generation of the predicate part of corresponds to the total number of membership func-
fuzzy rules (Create rule base); tions involved in the linguistic variables describing the
GA2 for optimization of the consequence part of corresponding input signals. The number of neurons of
fuzzy rules (Optimize rules); the third layer is given by the set of fuzzy rules involved
GA3, which represent the algorithm of readjustment in a given knowledge base. To choose a particular rule,
of the parameters of linguistic variables for a more it is necessary to choose the corresponding neuron of

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 757

Fig. 6. The editor of linguistic variables.

the third layer. The chosen rule can be further changed s-function of Simulink. To simulate fuzzy inference
and appended. (without using Simulink models), the corresponding
mex-file was prepared, which allows one to obtain the
In the bottom part of the main menu of the opti- results of fuzzy inference with the help of the command
mizer, the window for the output of system messages is line and executed scripts of Matlab. The program is
located, in which the parameters of algorithms and all compatible with Matlab 6.1 and subsequent versions.
actions made by the user are copied. The constantly
updated result of fuzzy inference is output together Since the main chain in the technology for designing
with the approximating training signal. Any actions intelligent control systems is the stage of designing the
aimed at a change of parameters of the designed knowl- corresponding knowledge base [1], the design of robust
edge base results in updating the approximation results. knowledge bases under the types of unpredicted control
Thus, the user can visually control the effect of modifi- situations specified above allows one to establish in a
cation of parameters of the knowledge base on the general form the correspondence between the condi-
result of the approximation. tions of functioning of the controlled plant and the
required robustness level of the intelligent control sys-
In the design of this system, it was initially planned tem. Consider the results of simulation of robust struc-
to use it together with Matlab, which allows one to flex- tures of intelligent control systems with efficient appli-
ibly compute the values of the fitness functions of cation of the soft computing optimizer.
genetic algorithms. Note that, together with the training
signal, it is possible to apply the results of numerical Remark 5. In Section 4, we described a methodol-
integration of models of the controlled plant executed ogy for designing robust knowledge bases and the cor-
in the Simulink environment controlled by a fuzzy con- responding software tools in the form of a soft comput-
troller with the synthesized soft computing optimizer. ing optimizer based on soft computing, which allows
An approach that allows one to compute the fitness one to solve the problem posed within the framework of
function in Matlab with the subsequent transfer of the processes of learning and adaptation. In what follows,
results to the genetic algorithm of the optimizer was we consider particular examples of application of the
developed. For this purpose, the corresponding library soft computing optimizer in the problems of testing and
of units of the Simulink environment was designed. evaluating the levels of structural robustness of the
This library supports the loading of the knowledge base designed intelligent control system based on the joint
and fuzzy inference (in the simulation mode), as well as technique of stochastic and fuzzy simulation. As simu-
the communication with the optimizer (in the optimiza- lation objects, we chose benchmarks that allow us to
tion mode). The unit of fuzzy inference for Simulink demonstrate clearly the efficiency and advantage of the
was written in C++, in the form of the corresponding developed tools for designing the soft computing opti-

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

758 LITVINTSEVA et al.

Fig. 7. The editor of the base of fuzzy rules.

mizer. The employed models of the controlled plant the stochastic process that allows one to investigate
possess both local and global dynamic instabilities, individual parameters of dynamic fuzzy systems.
high sensitivity to variation of the initial conditions, Selective trajectories should meet these requirements if
parameters of the structure of the controlled object, and their density function of the probability distribution is
random parametric, internal, and external perturba- known. Stochastic processes with a required density
tions. We present the results of simulation and practical function of probability distribution are simulated by the
recommendations for using them in the problems of method of nonlinear forming filters.
designing robust intelligent control systems. The meth- In this section, we use the methodology of designing
odology of stochastic simulation is described in short in the structures of intelligent control systems functioning
the Appendix. in the external environment under the presence of sto-
chastic processes having the same autocorrelation func-
tion and different distribution functions of the probabil-
5. A SYSTEM OF STOCHASTIC FUZZY ity density. The method of nonlinear forming filters for
SIMULATION OF ROBUST INTELLIGENT describing stochastic processes with a required density
CONTROL SYSTEMS function for the probability distribution based on the
Fuzzy simulation of robust knowledge bases with Fokker–Planck–Kolmogorov equations is described in
the soft computing optimizer is based on the process of the Appendix. This approach allows us to develop a
extraction of valuable information by simulation and generalized methodology for investigating the robust-
investigation of individual (statistically represented) ness of intelligent control systems based on stochastic
informative trajectories describing the behavior of the fuzzy simulation.
controlled plant and a conventional PID controller Figure 8 presents the generalized structure of the
under the effect of stochastic processes. Within the system of stochastic fuzzy simulation, which was
scope of correlation theory, stochastic processes, which applied for evaluating the robustness and limiting capa-
are different in their statistical nature (i.e., having dif- bilities of the structures of intelligent control systems
ferent density functions of probability distribution), can with specifying the main factors that affect the sensitiv-
be indistinguishable in their correlation properties. The ity and reliability of control. The efficiency of applica-
density function of probability distribution is the com- tion of the soft computing optimizer is demonstrated by
plete statistical characteristics of stochastic processes. particular typical examples of models of controlled
Therefore, the output process of the forming filter sim- plants, the so-called benchmarks.
ulating the external environment must be represented In particular, the investigated models of physical
by the informatively significant selective trajectory of controlled plants and their functioning environment are

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 759

Designed KB
SCO

FC

Adjustment of gains
in real time
(Y) PID Controlled (X)
Reference
signal + controller plant
– + +

m(t)

Noise
Z –1

Time delay
Control situation

Fig. 8. The block diagram of stochastic fuzzy simulation with unpredicted control situations.

characterized by the following specific features typical trol goal), we can simulate unpredicted control situa-
of real dynamic controlled plants: tions and evaluate the sensitivity and the robustness
they have local and global dynamic instability with level of the designed intelligent control system.
respect to the generalized coordinates;
In this section, we present the results of simulating
they have essentially nonlinear cross constraints the robust control laws for intelligent fuzzy PID con-
(stochastic nonlinearities) in the generalized dynamic trollers by complex essentially nonlinear dynamic con-
coordinates, which mutually affect (antagonistically)
the dynamic, stability, and controllability of the con- trolled plants. To demonstrate the capabilities of simu-
trolled plant; lation of the processes of intelligent control of a
dynamic controlled plant and the conditions of func-
they operate under unpredicted control situations. tioning, the results of simulation of the following three
As unpredicted control situations, we consider four typical controlled plants (benchmarks) are considered:
control models under the conditions of uncertainty of (1) a nonlinear oscillator with essential dissipation and
the source information: (1) with statistical information local dynamic instability; (2) an inverted pendulum
about the external and parametric random perturbations mounted on a moving cart and with global dynamic
variable in time (selective trajectories of stochastic pro-
cesses with density functions of probability distribution instability; and (3) an essentially nonlinear oscillator
depending on time); (2) with uncertainty of information with local and global dynamic instability in cross con-
about the variation of parameters of the structure of the straints of the generalized coordinates of the controlled
controlled plant; (3) under the presence of random plant.
delay time in the loops of control and measurement sys- These oscillators are of independent interest for
tems; and (4) when the control (reference signal) goals problems in robotics and mechanics (e.g., a strobo-
are changed.
scopic manipulator robot with complex behavior
The developed model of the intelligent control sys- dynamics and considerable dissipation) and allow one
tem and controlled plant was simulated in the Mat- to compare our results with the results obtained by
lab/Simulink system presented in Fig. 9. As typical ran- methods based on a fuzzy neural network [1].
dom noise, three types of stochastic processes with the
corresponding density functions of probability distribu- Remark 6. In view of the large amount of the simu-
tion were simulated. lation results and the limited admissible size of the
Figure 10 presents the form of the density functions paper, we consider the third version of an oscillator
of probability distribution and the simulation results of containing all the qualitative specific features of the two
the output stochastic processes from the corresponding types of oscillators listed above.
forming filters. Varying the structure of the forming fil-
ters, the parameters in the models of the controlled Example 2. A swing. Figure 11 shows a physical
plant, the delay time in the channel for measuring the swing. The motion equations of the pendulum are repre-
control error, and the form of the reference signal (con- sented as

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

760 LITVINTSEVA et al.

Fuzzy controller
Input PID gains
SCoptimizer based
FIS Output
Weights
simoutK 12:34 time
SCoptimizer S-Function1
To Workspace3 Digital Clock To Workspace8
Control error

Ground
Model of the CP simoutX
simoutRef
To Workspace4 simoutU Out1 To Workspace1

To Workspace6 In1
+ PID Out2
–

K ++ Pendulum Cart
Derivative Controlled U simoutZ
Ref Error
du/dt

PID controller Saturation To Workspace6

X dSo
simoutErr
PID controller z
To Workspace7
Integrator
1/s 1
Noise
Noise_Gain
g3_1t120.mat simoutSo

To Workspace2
Noise
Forming simoutN
Saturation filter To Workspace5 z

f (u) Fon Selector1

Unit Delay
angle
1/z

Time delay Selector2

in the measurement system

Fig. 9. A Matlab/Simulink Model of the control system.

⎧ The motion of the system described by Eqs. (5.1)

l˙
⎪ θ̇˙ + 2-θ̇ g
∫
represents a globally instable (with respect to the axis l)
+ --- sin θ = k 'p e θ + k 'd ė θ + k 'i e θ dt + ξ 1 ( t )
⎪ l l and locally instable (with respect to the axis θ) dynamic
⎪ system (Fig. 12). Note that, in model (5.1), there are
⎨ l̇˙+ 2kl˙ – lθ̇ 2 – g cos θ (5.1) cross nonlinear constraints that affect the local instabil-
⎪ ity with respect to the generalized coordinate θ. Thus,
⎪ --- 1
⎩
∫
⎪ = m- (k p e l + k d ė l + k i e l dt + ξ 2 ( t )), (5.1) involves local and global instability with respect
to the generalized coordinates of the controlled plant
and is a generalization of the previous models. Here and
where ξ1(t) and ξ2(t) are the corresponding stochastic in what follows, in Figs. 13–23, we use the following
actions described by various density functions of the notation: X: angle (θ = x) is an angle; Y: Length (l = y)
probability distribution. is a length; Time (S) is the time; SCO for a soft comput-
ing optimizer; and Refsignal for a reference signal.
The thermodynamic equations of the entropy pro- Consider the following problem of feedback control:
duction rate in the controlled plant and the PID control- in the presence of Rayleigh stochastic noise (see Fig. 10,
ler have the form position 2) acting on the controlled plant along the axis θ
(the maximum amplitude of noise is ten), and under
dS θ l˙ dS
-------- = 2-θ̇θ; -------l = 2k i l˙; Gauss stochastic noise along the axis l (the maximum
dt l dt amplitude of noise is four), steer the pendulum from the
⎧
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎩

Controlled plant initial position to a given position (control goal) and hold
(5.2) the motion of the controlled plant in this state θref = 0.4
dS PID ( θ ) 2 dS PID ( l ) 2 and lref = 3.5.
------------------ = k 'd ė θ ; ----------------
- = k d ė l
dt dt Consider the control problem under the following
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎩

PID controller values of parameters: m = 1 and k = 1 and the initial

and are considered together with motion equations of conditions [x0 = 2.5, y0 = 25} and [ ẋ 0 = 0, ẏ 0 = 0.01].
the controlled plant (5.1). The structure of the intelligent control system is a fuzzy

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 761

p(ε)
0.040 4
2
0.035 1
0
–2
0.030
–4
0.025
2 4
0.020 2
0
0.015 –2
–4
0.010 4
3
2
0.005 0
–2
0 –4
–5 –4 –3 –2 –1 0 1 2 3 4 5 0 10 20 30 40 50 60
ζ(t), |N|

Fig. 10. The form of the density function of the probability distribution and the results of simulation of output stochastic processes
from the corresponding forming filters: (1) Gaussian; (2) Rayleigh; (3) uniformly distributed.

controller for two PID controllers along the axes θ and l. which an optimal training signal was chosen, and com-
In this case, we consider the problem of coordinated con- pare the results of simulation for the three control cases.
trol for the gains of two PID controllers, which is of inde- Control problem with new conditions:
pendent interest for control theory and systems [2, 3, 8].
new initial position of the controlled object [–0.52,
Figures 13–17 show the results of simulation of the 2.5] [0.01, 0];
system motion in the three cases of control: with the new reference signals: θref = 0.78; lref = 5;
help of two classical PID controllers (with gains K = {6, other noises acting the controlled plant;
6, 6}); based on conventional soft computing (with the
help of the structure of a fuzzy neural network chosen Gaussian noise along the axis θ (the maximum
manually; and the method for designing knowledge amplitude of the noise is 1.5);
bases with the help of the error-back-propagation Rayleigh noise along the axis l (the maximum
method; and with application of tools of an soft com- amplitude of noise is equal to one).
puting optimizer developed in [3, 18, 21]. Therefore,
the intelligent control system designed based on the
soft computing optimizer is more efficient than the
intelligent control system obtained with the help of
conventional soft computing using a fuzzy neural net-
work, and more efficient than conventional PID con-
trollers. From the point of view of the processes of opti- l
mization of the structures of automatic control systems
according to the control quality criteria, such as mini-
mum of control error; minimum entropy production in
the controlled plant and in the automatic control system
(minimum heat loss, loss of useful work and energy), θ
the developed structure of the intelligent control system
is optimal and has a minimum complexity.
m
Investigation of robustness of the intelligent control
system. Consider the robustness property of the
designed knowledge bases for the three control cases
specified above. Let us formulate a control problem
with new conditions (unpredicted control situation),
which are different from the initial control problem for Fig. 11. A physical swing.

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

762 LITVINTSEVA et al.

X: angle Theta
0.3 0.7
Free motion PID
0.2 0.6 FNN
SCO
0.1 0.5
0 0.4
–0.1 0.3
Y: length 0.2
250 L
200
150 6
PID
100 5 FNN
50 Unstable motion SCO
4
0 5 10 15 20 25 30 35 40 45 50
Time, s 3
2
Fig. 12. A swing: free motion of the system. The coordinate 0 5 10 15 20 25 30
Y (length) shows globally unstable motion. Time, s

Fig. 13. A swing. The motion of the system under the sto-
chastic action with three control types (PID, FNN, and
SCO).
Sp
20
15 U(t) along x
10 3 PID
PID 2 FNN
5 FNN SCO
SCO 1
0
Sc 0
50 –1
40 U(t) along y
30 10
20 PID 0
10 FNN
SCO –10
0 5 10 15 20 25 30 –20
Time, s –30
0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
Fig. 14. A physical swing. Entropy production in the con- Time, s
trolled plant and in the control system (PID, FNN, and
SCO). Fig. 15. A physical swing. The control force along the coor-
dinate axes θ(x) and l(y).

Figures 18–20 show the results of simulation of the

system motion in the three control cases under the new obtained with application of the soft computing opti-
conditions. mizer is robust and efficient from the point of view of a
In Figs. 21–23, the control force and the gains of the given control quality criterion.
fuzzy PID controller based on the conventional method
of soft computing presented by a fuzzy neural network
are zero. This means that the knowledge base of the 6. EXAMPLES OF PRACTICAL APPLICATION
fuzzy controller designed with the help of a fuzzy neu- OF THE SOFT COMPUTING OPTIMIZER
ral network is not robust; i.e., for the new initial condi- IN ACTUAL CONTROLLED OBJECTS
tions, there are no activated rules in the knowledge
base. By the simulation results, the designed knowl- The tools of the soft computing optimizer based on
edge base based on a fuzzy neural network for the fuzzy soft computing were employed in designing robust
controller in the intelligent control system cannot cope intelligent control systems of industrial controlled
with the control situation. Therefore, the knowledge plants were experimentally tested on actual controlled
base of the intelligent control system designed with the objects, such as an inverted physical pendulum without
help of conventional soft computing based on a fuzzy a mathematical model, semi-active car suspension [23],
neural network is not robust. On the contrary, the navigation control for a robot–cyclist [24], control of a
knowledge base of the intelligent control system combustion engine [25], etc.

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 763

Kp gain x
15 PID
FNN
10 SCO

5
0
Kd gain x
20
10
0
–10
Ki gain x
15
10
5
0
0 1 2 3 4 5 6 7 8 9 10
Time, s

Fig. 16. Control laws force along the coordinate axes θ (PID, FNN, and SCO).

Kp gain y
15 PID
FNN
10 SCO
5
0
Kd gain y
15
10
5
0
Ki gain y
20
10
0
–10
0 1 2 3 4 5 6 7 8 9 10
Time, s

Fig. 17. Control laws force along the coordinate axes l (PID, FNN, and SCO).

The conducted tests have shown a high efficiency of from the viewpoint of the control quality criteria
the operation of the soft computing optimizer both on such as the minimum control error, the minimum
the controlled plant with a model and with the use of entropy production in the controlled plant and control
experimental data obtained with a physical model of system (i.e., the minimum heat loss and loss of useful
the controlled plant. The results of simulation and
industrial exploitation have demonstrated that: work and energy), as well as with due account of the
the designed knowledge base of the fuzzy controller, minimum of control force the developed structure of
which controls the motion of the controlled plant, is the intelligent control system, is more efficient than the
robust; conventional PID controllers and fuzzy controllers

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

764 LITVINTSEVA et al.

X Control error along x axis

1.5 2.0
PID
1.0 1.5 FNN
0.5 SCO
1.0
0 PID
FNN 0.5
–0.5 SCO 0
Refsignal
–1.0 –0.5
Y Control error along y axis
200 50
150 0
PID
100 PID –50 FNN
FNN SCO
50 SCO –100
Refsignal
–150
0 5 10 15 20 25 30 –200
Time, s 0 5 10 15 20 25 30
Time, s
Fig. 18. A physical swing. The motion of the system under
the stochastic action with three control types (PID, FNN, and Fig. 19. A physical swing. A control error. A new control sit-
SCO). A new control situation. uation.

Sp U(t)along x
2000 15
PID
1500 10 FNN
PID SCO
1000 5
FNN
500 SCO 0
–5
0 5 10 15 20 25 30 0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
Sc U(t)along y
100 20
80 PID
10 FNN
60 PID SCO
40 FNN 0
20 SCO –10
0 5 10 15 20 25 30 –20
0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
Fig. 20. A physical swing. Entropy production in the con- Fig. 21. A physical swing. The control force along the coor-
trolled plant and in the control system. A new control situa- dinate axes θ(x) and l(y). A new control situation.
tion.

designed based on the conventional tools of soft com- ing and adaptation, implementing the first stage of
puting of the fuzzy neural network type; information technology of designing robust intelligent
the developed methods has its limitations in achieve- control systems.
ment of the required robustness level, which makes us
introduce new methods of simulation and computing
(quantum computing and quantum fuzzy inference) to 7. CONSTRAINTS AND PROSPECTS
the tools of the soft computing optimizer [7]. OF DEVELOPMENT OF SOFTWARE TOOLS
OF THE SOFT COMPUTING OPTIMIZER
Based on the proposed tools of the soft computing
optimizer, the topical problems of forming knowledge As was mentioned, using the integration with Mat-
bases for designing robust fuzzy controllers were con- lab, the soft computing optimizer gives the opportunity
sidered, e.g., the problem of coordinated control of to introduce additional control quality criteria not
gains of two PID controllers, which are of independent changing the source code of software units of the sys-
interest for control theory and systems. The employed tem. In view of the advantages listed above, the soft
tools of the soft computing optimizer allow one to computing optimizer provides a background for the
implement simultaneously the process of designing first stage of technology of designing intelligent control
robust knowledge bases based on algorithms of learn- systems [3, 20, 21] (Fig. 3).

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 765

Kp gain x
10

5 PID
FNN
SCO
0
Kd gain x
10

0
Ki gain x
10

0 5 10 15 20 25 30
Time, s

Fig. 22. Control laws force along the coordinate axes θ(x). A new control situation.

Kp gain y
10

5 PID
FNN
SCO
0
Kd gain y
10

0
Ki gain y
10

0 5 10 15 20 25 30
Time, s

Fig. 23. Control laws force along the coordinate axes l. A new control situation.

Remark 7. The large capabilities of the tools of the limitations in achievement of the required robustness
soft computing optimizer in designing robust knowl- level.
edge bases of intelligent control systems for a wide Figures 24 and 25 present the results of simulating
class of control situations have been investigated and the “cart-pole” and “swing” systems in unpredicted
demonstrated. However, the proposed method also has control situations. For the “cart-pole” system, the new

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

766 LITVINTSEVA et al.

X: cart position X (a)

400 100
New situation with delay = 0.008 : FC
300 PID 50
200
0
100 S2a: PID average
–50 FC1
0 FC2
Refsignal
–100 –100
Z: cart position Y
4 5
2 4
0 3
2
–2
1
–4
0 5 10 15 20 25 30 35 40 45 50 0
Time, s Sp (b)
100
Fig. 24. Unstable dynamic behavior of the CP in a new 80
unpredicted control situation (the time delay is enlarged to 60
0.008 s). 40 S2a: Total entropy of plant: PID average
FC1
20 FC2

delay time in the measurement channel was 0.008 s 0

(under the delay time 0.001 and 0.002 s the system is Sc
stable and robust) and resulted in loss of control for the 2000
fuzzy and conventional PID controllers. 1500
For the “swing” system, the unpredicted situation 1000
S2a: Total entropy of 2 Controllers: PID average
for the initial parameters of the system m = 0.5, k = 0.4, 500 FC1
FC2
c = 2; and [θ0 = 0.25, l0 = 0.5] [ θ̇ 0 = 0, l˙0 = 0.01] had
the following characteristics: Gaussian noise along the 0 2 4 6 8 10 12 14 16 18 20
axis x, and the time delay in the measurement channel Time, s
was 0.075 s. The Rayleigh noise along the axis y and
the time delay in the measurement channel were 0.04s, Fig. 25. The dynamic behavior of the CP in an unpredicted
control situation (the description in the text): (a) loss of con-
and the reference signals along the axis x and axis y are trollability of the PID controller; (b) entropy characteristics
equal to zero and two, respectively. Even without a lim- of the CP and controllers.
itation on the control force, the conventional PID con-
troller with averaged optimal parameters cannot cope
with the unpredicted control situation. set with the help of the quantum superposition operator
[4, 7]. This type of computing is quantum computing.
Remark 8. In Figs. 24 and 25, we use the following
notation: X is the angle of pendulum deflection; Z is the In this case, the quantum optimizer that applies
cart position; Sp is the entropy production in the con- methods of quantum computing uses particular knowl-
trolled plant; Sc is the entropy production in the control- edge bases designed based on the soft computing opti-
ler; and Kp, KD, and KI are the gains of the PID control-
ler. These and other simulation results lead to the neces- mizer as the source data. This means that the soft com-
sity to develop and introduce new methods and puting optimizer provides a background for implement-
computing (quantum computing and quantum fuzzy ing the second stage of technique using quantum
inference [7, 13] to the tools of soft computing opti- computing. For this purpose, an interface that can be
mizer. inserted by a program with a unit of quantum comput-
The constraints arise from the physical nature of the ing is incorporated in the architecture of the soft com-
algorithm of random search in the structure of genetic puting optimizer (Fig. 26).
algorithm. The proposed soft computing optimizer is Thus, the promising information technology of
able to work on a single space of solutions (because of
limit capabilities of the genetic algorithm). For global designing the structures of robust intelligent control
optimization on the set of spaces of solutions, it is systems is two-level. The first level is based on the soft
required to pass to a new type of computing that allows computing optimizer, and the second level is based on
one to join the sets of spaces of solutions into a single quantum computing.

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 767

Matlab
Soft computing optimizer QC
Export prospective system

Real-time
system

Choice of a training signal Choice of a model Choice

of fuzzy inference of a training
signal

Other Sugeno
order 0
GA VC classical or 1 Mamdani Tsukamoto Fitness
methods Table function
of inputs
outputs

Choice of Choice of the class

Rules Membership inference of membership Model Controlled
functions type
functions in Matlab plant

Rules and
membership
functions
FNN

Fig. 26. The block diagram of the CSO.

CONCLUSIONS measurement system without loss of accuracy and con-

trol quality [22].
The methods for development and the technology The robust intelligent control system obtained based
for designing intelligent control systems based on the on this approach requires minimum source data on both
soft computing optimizer presented in this paper allow the behavior of the controlled plant and the external
one to implement the principle of designing an optimal perturbations [3, 20].
intelligent control system with a maximum reliability
and controllability level of a complex controlled plant
under conditions of uncertainty in the source data, and APPENDIX
in the presence of stochastic noises of various physical The system for simulating stochastic processes with
and statistical characters. The knowledge bases formed a required density function of probability distribution
with the help of a soft computing optimizer produce by the method of nonlinear forming filters. In this
robust control laws for the gains of conventional PID appendix, we describe the methodology of designing
controllers for a wide range of external perturbations the structures of forming filters of stochastic processes
and are maximally insensitive to random variations of that have the same autocorrelation function and differ-
the structure of the controlled plant. ent functions of the probability distribution density. The
results of statistical analysis of actual stochastic pro-
The robustness of control laws is achieved by intro- cesses have shown that the typical autocorrelation func-
ducing a vector fitness function for the genetic algo- tions describing these processes are
rithm, whose one component describes the physical
principle of minimum production of generalized R ( τ ) = B ( 0 ) exp { –α τ }; (A.1)
entropy both in the controlled plant and the control sys- R ( τ ) = B ( 0 ) exp { –α τ } cos βτ; (A.2)
tem, and the other components describe conventional
control objective functionals such as minimum control α
error, etc. The approach based on the soft computing R ( τ ) = B ( 0 ) exp { –α τ } cos βτ + --- sin ( β τ ) , (A.3)
β
optimizer for designing robust intelligent control sys-
tems allows one to design an optimal intelligent control where α and β are the values of coefficients determined
system with a maximum reliability and controllability from the experimental data. The presented autocorrela-
level for the set of dynamic systems under the presence tion functions and their parameters are used in the prac-
of uncertainty in the source data [8]; to reduce the num- tice of stochastic simulation of stochastic processes
ber of sensors both in the control channel and in the based on forming filters.

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

768 LITVINTSEVA et al.

In the framework of correlation theory, stochastic density function of probability distribution p(x). It is
processes that are different in their nature (i.e., having known that the density function of probability distribu-
different density functions of probability distribution) tion for a given class of dynamic systems can be deter-
may be indiscernible in their correlation properties. The mined by solving the Fokker-Planck–Kolmogorov
density function of probability distribution is a com- equation [26, 28], which can be obtained in the station-
plete statistical characteristic of stochastic processes. ary case from (A.5) as follows:
Consider the methodology of designing the structures
of forming filters of stochastic processes by the exam- d d⎧ 1d ⎫
------ G = – ------ ⎨ αxp ( x ) + --- ------ [ D ( x ) p ( x ) ] ⎬ = 0, (A.8)
2
ple of an autocorrelation function (A.1) with various
types of density functions of probability distribution. dx dx ⎩ 2 dx ⎭
Let us take a stationary stochastic process X(t)
defined on the interval [xl, xr] whose boundaries may be where G is known as the probability flow. Since the sto-
both bounded and unbounded. Without loss of general- chastic process X(t) is defined on the interval [xl, xr], the
ity, we assume that X(t) has a zero mean value. Then, probability plow G must satisfy two boundary condi-
we have xl < 0 and xr > 0. Knowing the density function tions for x = xl and x = xr.
of probability distribution p(x) and the spectral density In this one-dimensional case, the flow G exists and
ΦXX(ω) (as the Fourier transform of the autocorrelation is stationary; i.e., the density of the probability flow is
function RXX(τ) of the stochastic process X(t)), we can
simulate the stochastic process X(t) with the help of a d
------ G = 0. In accordance with this condition, Eq. (A.8)
forming filter, whose structure and parameters are dx
determined in terms of the form and parameters of the takes the form
autocorrelation function and density function of the
probability distribution [26–28]. 1d
αxp ( x ) + --- ------ [ D ( x ) p ( x ) ] = 0.
2
(A.9)
Assume that the spectral density of the stochastic 2 dx
process has the form
The integration of Eq. (A.9) leads to the result
ασ
2
Φ XX ( ω ) = --------------------------, α > 0, (A.4) xr
π(ω + α )
2 2

∫
D ( x ) p ( x ) = –2α up ( u ) du + C,
2
(A.10)
where σ2 is the mean square deviation of X(t). If X(t) is xl
a Markovian diffusion process, then it can be consid-
ered as a solution to the following stochastic equation where C is the integration constant.
(in the Ito sense):
To determine C, we consider two cases. In the first
dX = –αXdt + D ( X )dB ( t ), (A.5) case, if either xl = –∞ or xr = ∞, or both cases are true,
where B(t) is a unit Wiener process, the coefficients –αX then p(x) = 0 on these boundary conditions, and
and D(X) are known as the shift and diffusion coeffi- Eq. (A.10) implies that C = 0. In the second case, if xl
cients of the stochastic process X(t), respectively. We and xr are finite, then the shift coefficient –αxl for the
multiply the left and right sides of (A.5) by X(t – τ) and right boundary condition is negative, while, for the left
average over the ensemble of random trajectories. As a boundary condition, it is positive.
result, we obtain This means that the averaged probability flows for
dR ( τ ) these two boundary conditions have different direc-
-------------- = –αR ( τ ), (A.6) tions. However, the existence of a stationary probability
dτ distribution density gives an additional condition that
where R(τ) is the autocorrelation function of the sto- states that, for these two boundary conditions, the dif-
chastic process X(t); i.e., R(τ) = E[X(t – τ)X(t)]. Equa- fusion coefficient is D2(xl) = D2(xr) = 0. This condition
tion (A.6) has the following solution is valid only if C = 0. As a result, we have
R ( τ ) = A exp ( –α τ ), (A.7) xr
2α
∫
D ( x ) = – ----------- up ( u ) du.
2
where A is the corresponding normalizing factor. (A.11)
p( x )
Taking A = σ2, Eqs. (A.4) and (A.7) can be related xl
by the Fourier transform.
Thus, Eq. (A.5) reproduces the stochastic process The function D2(x) computed from Eq. (A.11) is nonne-
X(t) with the required spectral density (A.4). Note that gative since p(x) ≥ 0 and the mean value of X(t) is zero.
the form of the diffusion coefficient D(X) does not Thus, the stochastic process X(t) is reproduced as a
affect the form of the spectral density. solution to Eq. (A.5) with the diffusion coefficient D(x)
Let us define the form of the diffusion coefficient determined from (A.11) with the required density function
D(X) so that the random process X(t) has the required of probability distribution p(x) and spectral density (A.4).

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 769

Table 3. The structure of forming filters for typical density functions of probability distribution p(x)

Correlation function Probability distribution function Structure of the forming filter

2 –α τ
Ry(τ) = σ e Gaussian ẏ + αy = σ ξ ( t )
2

2
2 –α τ α σ 2 2
Ry(τ) = σ e Uniform ẏ + --- y = ---------- α ( ∆ – y )ξ ( t )
2 2π

2
2α σ
ẏ + αy ± ------- = ---------- ------- ⎛ y + ---⎞ ξ ( t )
2 –α τ 2α 2
Ry(τ) = σ e Rayleigh
⎝
γ 2π γ γ⎠

α
ẏ + αy + -------------------- ( b 2 x + b 1 )
a 1 + 2b 2
2 –α τ
Ry(τ) = σ e Pierson
σ2 2α 2
= ---------- -------------------- ( b 2 y + b 1 y + b 0 )ξ ( t )
2π a 1 + 2b 2

Table 4. The structure of forming filters of stochastic processes with autocorrelation function (A.3) and typical density func-
tions of probability distribution

Correlation function Density function of probability distribution Structure of the forming filter

2D Gaussian 2 2
ẏ˙ + 2αẏ + ( α + ω )y
1 y1 – µ1 2 y2 – µ2 2
– --- ⎛ ⎛ ----------------⎞ + ⎛ ----------------⎞ ⎞
2 ⎝ ⎝ σ1 ⎠ ⎝ σ2 ⎠ ⎠ 2 2 2
1
p ( y 1, y 2 ) = ------------------- e = 2ασ ( α + ω )ξ ( t )
2πσ 1 σ 2

2D Uniform
a 11
ẏ 1 = --- a 11 y 1 + a 12 y 2 + ⎛ – ------- ( ∆ 1 – y 1 )⎞ ξ 1 ( t )
1 2
1 ⎝ 2π ⎠
p ( y 1, y 2 ) = --------------- 2
4∆ 1 ∆ 2
a 22
–∆ 1 < y 1 < ∆ 1 ẏ 2 = --- a 22 y 2 + a 21 y 1 + ⎛ – ------- ( ∆ 2 – y 2 )⎞ ξ 2 ( t )
1 2
2 ⎝ 2π ⎠
–∆ 2 < y 2 < ∆ 2

R(τ) 2D Hyperbolic 2
2a 11
2 –α τ α p ( y 1, y 2 ) = ρ ( λ ) = C 1 ( λ + b )
–δ ẏ 1 = a 11 y 1 + a 12 y 2 – ------------------2-
= σ e cos ωτ + ---- sin ω τ (δ – 1)
ω
b > 0; δ>1
1 2 a 12 2
× --- y 1 – ---------- y 2 + b y 1
1 2 a 12 2 2 2a 21
λ = --- y 1 – ---------- y 2
2 2a 21
2a 11 1 2 a 12 2
– --------------------------- --- y 1 – ---------- y 2 + b ξ 1 ( t )
2π ( δ – 1 ) 2 2a 21
2 3
2a 22 a 12
ẏ 2 = a 21 y 1 + a 22 y 2 + --------------------------2
3
a 21 ( δ – 1 )
1 2 a 12 2
× --- y 1 – ---------- y 2 + b y 2
2 2a 21
2a 22 a 12 1 2 a 12 2
+ ----------------------------------- --- y 1 – ---------- y 2 + b ξ 2 ( t )
2πa 21 ( δ – 1 ) 2 2a 21

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

770 LITVINTSEVA et al.

The stochastic differential equation (A.5) in the Ito Note that the spectral density for X(t) contains a delta-
form can be transformed to the corresponding Stra- function of the form (4/γ2)δ(ω) with the mean value 2/γ
tonovich form different from zero.
Example 5. Consider a family of density functions
1 d D ( X ) D( X )
2
Ẋ = –αX – --- ------------------- + -------------ξ ( t ), (A.12) of the probability distribution satisfying the equation
4 dX 2π
d
where ξ(t) is the Gaussian white noise with unit spectral ------ p ( x ) = J ( x ) p ( x ). (A.22)
dx
density. Equation (A.12) is more practical in simulation
of the selective trajectories of stochastic processes. Equation (A.22) can be integrated as
Consider certain illustrative examples that are of inde-
pendent interest for the theory and practice of stochas- ∫
p ( x ) = C 1 exp ( J ( x ) d x ), (A.23)
tic simulation of stochastic processes [26–28]. where C1 is a normalizing constant. Then, we have
X(t) Suppose that X(t) is a stationary stochastic pro-
∫
D ( x ) = –2α exp [ –J ( x ) ] x exp [ J ( x ) ] d x.
2
cess with a uniform density function of probability dis- (A.24)
tribution (Fig. 10, position 3).
Note the following special cases. Admit that
1
p ( x ) = ------- , –∆ ≤ x ≤ ∆. (A.13)
J ( x ) = –γ x – δx , –∞ < x < ∞,
2 4
2∆ (A.25)
Substituting (A.13) into (A.11), we obtain where γ is an arbitrary constant if δ > 0. Substituting
(A.25) into (A.11), we obtain
D ( x ) = α ( ∆ – x ).
2 2 2
(A.14)
D ( x)
2
In this case, the Ito stochastic differential equation has (A.26)
α γ 2 γ⎞
= --- π/δ exp δ ⎛ x + ------⎞ erfc ⎛
the form
δ x + ------ ,
2 2
2 ⎝ 2δ⎠ ⎝ 2δ⎠
dX = –αXdt + α ( ∆ – X )dB ( t ).
2 2
(A.15)
where erfc(y) is a special error function defined as
It is worth noting that the family of stochastic processes
∞
can be obtained from the following generalized form of
2 –t2
Eq. (A.15) erfc ( y ) = ------- e dt.
π ∫ (A.27)
y
dX = –αXdt + αβ ( ∆ – X )dB ( t ).
2 2
(A.16)
The case γ < 0 and δ > 0 corresponds to a bimodal dis-
The equations have different forms, but the same spec- tribution, and the case γ > 0 and δ = 0 corresponds to a
tral density (A.4). Gaussian distribution. The Pierson family of probabil-
Example 4. Admit that X(t) has a Rayleigh density ity distributions is represented in the form
function of probability distribution a1 x + a0
J ( x ) = -----------------------------------
-, (A.28)
p ( x ) = γ x exp ( –γx ), γ > 0, 0 ≤ x ≤ ∞. (A.17)
2 2
b2 x + b1 x + b0
Let us transform X(t) to the following form Y(t) = X(t) – and, in the special case for a0 + b1 = 0, we have
2/γ with the probability distribution density:
2α
D ( x ) = – ---------------- ( b 2 x + b 1 x + b 0 ).
2 2
p ( y ) = γ ( γy + 2 ) exp ( –γy + 2 ), –2/ γ ≤ y ≤ ∞ . (A.18) (A.29)
a1 + b2
Equation (A.11) implies that
Thus, the structure of forming filters with autocorrela-
tion function (A.1) and different density functions of
D ( y ) = ------- ⎛ y + ---⎞ .
2 2α 2
(A.19)
γ ⎝ γ⎠ the probability distribution can be described by typi-
cal structures of forming filters represented in Table 3.
The Ito stochastic differential equation for Y(t) has the By a similar method, the structures of forming filters
form for stochastic processes with autocorrelation functi-
ons (A.2) and (A.3) are determined.
2α ⎛
------- Y + ---⎞ dB ( t )
2 Table 4 presents the structures of forming filters for
dY = –αYdt + (A.20)
γ ⎝ γ⎠ autocorrelation function (A.3) for the typical density
functions of the probability distribution in the two-
vand the corresponding Stratonovich equation for X(t) dimensional case.
is Figure 10 presents examples of application of the
methodology for stochastic simulation of stochastic
α
Ẋ = –αX + ------- + ⎛ ------ X⎞ ξ ( t ).
3α processes [26–28] with the required density function of
(A.21)
2γ ⎝ πγ ⎠ the probability distribution. Other methods for simulat-

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006

 DESIGN OF ROBUST KNOWLEDGE BASES OF FUZZY CONTROLLERS 771

ing stochastic processes and guaranteed estimates of 15. A. A. Krasovskii, “Dangerous” Entropy Instability of
stochastic stability of schemes of numerical integration ACS, Dokl. Ross. Akad. Nauk 327 (6) (2000).
were described in [29–33] and were taken into account 16. M. Bologna, P. Grigoloni, M. Karagiorgios, et al., “L.
in the simulation of the considered stochastic nonlinear Trajectory Versus Probability Density Entropy,” Phys.
models of the controlled plant. Rev. 64E (1) (2001).
17. U. Seifert, “Entropy Production along a Stochastic Tra-
jectory and an Integral Fluctuation Theorem,” Phys. Rev.
REFERENCES Lett. 95 (4) (2005).
1. I. Kuravaki, L. V. Litvintseva, K. Takakhashi, et al., 18. S. A. Panfilov, L. V. Litvintseva, S. S. Ul’yanov, et al.,
“Design of Robust Knowledge Bases of Fuzzy Control- “Software Support of the Processes of Generation, Elicit-
lers for Intelligent Control of Substantially Nonlinear ing, and Design of Knowledge Bases of Robust Intelligent
Dynamic Systems, I, Application of Soft Computing,” Control Systems,” Program. Prod. Sist., No. 2 (2004).
Izv. Ross. Akad. Nauk, Teor. Sist. Upr., No. 4 (2004) 19. V. N. Zakharov and S. V. Ul’yanov, “Industrial Fuzzy
[Comp. Syst. Sci. 43 (3), 615–632 (2004)]. Control Systems and Fuzzy Controllers, I, Science-
2. B. N. Petrov, I. I. Gol’denblat, S. V. Ul’yanov, et al., Organizational and Technical–Economic Aspects,” Izv.
The Theory of Models in Control Processes: Informa- Akad. Nauk SSSR, Tekh. Kibern., No. 5 (1992).
tional Thermodynamic Aspects (Nauka, Moscow, 20. S. A. Panfilov, L. V. Litvintseva, S. V. Ulyanov, et al.,
1978) [in Russian]. “Soft Computing Optimizer of Intelligent Control Sys-
3. L. V. Litvintseva, K. Takahashi, S. S. Ulyanov, et al., tem Structures.” US patent, No. 20,050,119,986 (2005).
“Intelligent Robust Control Design Based on New Types 21. S. A. Panfilov, L. V. Litvintseva, S. V. Ulyanov, et al.,
of Computations,” Note del Polo Ricerca. Milana: Uni- “Soft Computing Optimizer for Intelligent Control Sys-
versita degli Studi di Milano Publ. 60 (2004). tems Design: the Structure and Applications,” J. Syst.,
Cybern. Informatics (USA) 1 (5) (2003).
4. B. N. Petrov, G. M. Ulanov, S. V. Ul’yanov, et al., Con-
trol Problems for Relativistic and Quantum Dynamic 22. S. V. Ulyanov, System for Intelligent Control Based on
Systems (Nauka, Moscow, 1982) [in Russian]. Soft Computing. US patent, No. 6 415272 (1998).
5. R. A. Aliev and S. V. Ul’yanov, “Fuzzy Models of Con- 23. S. V. Ulyanov, S. A. Panfilov, and T. Hagiwara, Intelli-
trol Processes and Systems,” Itog. Nauk. Tekh. (VIN- gent Mechatronic Control Suspension System Based on
ITI), Ser. Tekh. Kibern. 29 (1990); 32 (1991). Soft Computing. US patent, No. 6,701,236 (2001).
24. S. Fujii, H. Watanabe, S. A. Panfilov, et al., l. Intelligent
6. B. N. Petrov, V. S. Pugachev, S. V. Ulyanov, et al., “Infor- Robust Control System for Motorcycle Using Soft Com-
mational Foundations of Qualitative Theory of Control puting Optimizer. US patent, No. 20050197994 (2005).
Systems,” in Proceedings of 7th IFAC Congress, Hels-
inki, Finland, 1978, Vol. 3. 25. S. V. Ulyanov, S. Hashimoto, and M. Yamaguchi, System
for Intelligent Control of an Engine Based on Soft Com-
7. S. V. Ul’yanov, A. V. Yazenin, K. Takakhashi, et al., “Mod- puting. US patent, No. 6,609,060 (2001).
eling and Design of Intelligent Robust Control Systems
with the Use of Quantum and Soft Computing: Since 1965 26. S. V. Ulyanov, M. Feng, K. Yamafuji, et al., “ Stochastic
together with Our Teacher, B.N. Petrov,” in Abstracts of Analysis of Time-Variant Nonlinear Dynamic Systems.
Papers of the Conference on Control Theory Dedicated to Pts 1, 2: the Fokker–Planck–Kolmogorov Equation
the Memory of Academician B.N.Petrov, Moscow, Russia, Approach in Stochastic Mechanics,” Prob. Engin. Mech.
2003 (Inst. Probl. Upr., Moscow, 2003). 13 (3) (1998).
8. S. V. Ulyanov, Self-organized Control System. US patent, 27. G. Q. Cai and C. Wu, “Modeling of Bounded Stochastic
No. 6, 411, 944 (1997). Processes,” Prob. Engin. Mech. 19 (3) (2004).
28. N. A. Nikolaenko and S. V. Ul’yanov, Statistical Dynam-
9. S. N. Vasil’ev, A. K. Zherlov, E. A. Fedosov, et al., Intel- ics of Engineering Structures (Mashinostroenie, Mos-
ligent Control for Dynamic Systems (Fizmatgiz, Mos- cow, 1977) [in Russian].
cow, 2000) [in Russian].
29. D. J. Higham, “An Algorithmic Introduction to Numeri-
10. B. T. Polyak and P. S. Shcherbakov, Robust Stability and cal Simulation of Stochastic Differential Equations,”
Control (Nauka, Moscow, 2002) [in Russian]. SIAM Rev. 43 (3) (2001).
11. O. N. Granichin and B. T. Polyak, Randomized Estima- 30. K. Burrage, P. M. Burrage, and T. Tian, “Numerical
tion and Optimization Algorithms under Almost Arbi- Methods for Strong Solutions of Stochastic Differential
trary Noise (Nauka, Moscow, 2003) [in Russian]. Equations: An Overview,” Proc. R. Lond. Soc. 460A
12. S. V. Ulyanov, K. Yamafuji, I. Kurawaki, et al., “Compu- (3946) (2004).
tational Intelligence for Robust Control Algorithms of 31. X. Mao, “Numerical Solutions of Stochastic Functional
Complex Dynamic Systems with Minimum Entropy Differential Equations,” LMS J. Comput. Math. 6 (2)
Production, I,” J. Advanced Comput. Intell. 3 (2) (1999). (2003).
13. S. V. Ulyanov, “Quantum Soft Computing in Control 32. D. J. Higham, X. Mao, and A. M. Stuart, “Exponential
Process Design: Quantum Genetic Algorithms and Mean-Square Stability of Numerical Solutions to Sto-
Quantum Neural Network Approaches,” in Proceedings chastic Functional Differential Equations,” LMS J.
of WAC'2004: Soft Computing and Industrial Applica- Comput. Math. 6 (3) (2003).
tions, Spain, 2004, Vol. 17. 33. J. A. Acebron and R. Spigler, “Fast Simulations of Sto-
14. M. Perroud and A. Saucier, “Thermodynamics of Dissi- chastic Dynamical Systems,” J. Comput. Phys. 208 (2)
pative Systems,” Helvetica Physica 60 (8) (1987). (2005).

JOURNAL OF COMPUTER AND SYSTEMS SCIENCES INTERNATIONAL Vol. 45 No. 5 2006