You are on page 1of 16

Soft Comput (2011) 15:11451160

DOI 10.1007/s00500-010-0588-9

FOCUS

Optimization of interval type-2 fuzzy logic controllers


using evolutionary algorithms
O. Castillo P. Melin A. Alanis O. Montiel

R. Sepulveda

Published online: 12 March 2010


 Springer-Verlag 2010

Abstract A method for designing optimal interval type-2 1 Introduction


fuzzy logic controllers using evolutionary algorithms is
presented in this paper. Interval type-2 fuzzy controllers The design of intelligent controllers for non-linear plants is
can outperform conventional type-1 fuzzy controllers when a complicated problem due to the complex interactions
the problem has a high degree of uncertainty. However, between the plants components, the nonlinearities present
designing interval type-2 fuzzy controllers is more difficult in the plants model, and also due to the high number of
because there are more parameters involved. In this paper, design parameters involved. For this reason machine
interval type-2 fuzzy systems are approximated with the learning and cybernetics, concepts are used to automate the
average of two type-1 fuzzy systems, which has been design process with the goal of finding optimal controllers.
shown to give good results in control if the type-1 fuzzy In this sense, machine learning and cybernetics play a
systems can be obtained appropriately. An evolutionary fundamental role in automating the design of optimal
algorithm is applied to find the optimal interval type-2 controllers based on existing historical data. In addition, in
fuzzy system as mentioned above. The human evolutionary the real world there are many sources of uncertainty
model is applied for optimizing the interval type-2 fuzzy present, which complicates even more the problem. For
controller for a particular non-linear plant and results are this reason, fuzzy controllers can offer a way to manage
compared against an optimal type-1 fuzzy controller. A uncertainties present in real world processes.
comparative study of simulation results of the type-2 and Uncertainty affects decision-making and appears in a
type-1 fuzzy controllers, under different noise levels, is number of different forms. The concept of information is
also presented. Simulation results show that interval type-2 fully connected with the concept of uncertainty. The most
fuzzy controllers obtained with the evolutionary algorithm fundamental aspect of this connection is that uncertainty is
outperform type-1 fuzzy controllers. involved in any problem-solving situation because of some
information deficiency, which may be incomplete, impre-
Keywords Interval type-2 fuzzy logic  cise, fragmentary, not fully reliable, vague, contradictory,
Evolutionary algorithms  Fuzzy control or deficient in some other way (Klir and Yuan 1995). The
general framework of fuzzy reasoning allows handling
most of this uncertainty. Traditional fuzzy systems employ
type-1 fuzzy sets, which represent uncertainty by numbers
in the [0, 1] range. When something is uncertain, like a
measurement, it is difficult to determine its exact value,
O. Castillo (&)  P. Melin  A. Alanis and of course type-1 fuzzy sets make more sense than using
Tijuana, Institute of Technology, Tijuana, BC, Mexico crisp sets (Zadeh 1975a, b). However, it is not reasonable
e-mail: ocastillo@hafsamx.org to use an accurate membership function for something
uncertain, so in this case what we need is another type of
O. Montiel  R. Sepulveda
Center for Research in Digital Systems, IPN, Tijuana, BC, fuzzy sets, those which are able to handle these uncer-
Mexico tainties, the so called type-2 fuzzy sets (Mendel 2000). The

123
1146 O. Castillo et al.

effects of uncertainty in a system can be handled in a better 2008). For quantifying the errors, we utilized three widely
way by using type-2 fuzzy logic because it offers better used performance criteria, these are: integral of square
capabilities to cope with linguistic uncertainties by mod- error (ISE), integral of the absolute value of the error
eling vagueness and unreliability of information (Karnik (IAE), and integral of the time multiplied by the absolute
and Mendel 2001a, b; Mendel 1999; Liang and Mendel value of the error (ITAE) (Deshpande and Ash 1988). The
2000; Yager 1980). comparison is made under different noise values to mea-
Recently, we have seen the use of type-2 fuzzy sets in sure the effect of uncertainty (Martinez et al. 2009).
fuzzy logic systems to deal with uncertain information In a second set of tests, the parameters of the Gaussian
(Karnik et al. 2001; Mendel 1998). So we can find some membership functions (MFs) of the interval type-2 FLC
papers emphasizing on the implementation of a type-2 were obtained with the optimization method known as
fuzzy logic system (FLS) (Karnik et al. 1999); in others, it Human Evolutionary Model (HEM), which is described in
is explained how type-2 fuzzy sets let us model and Sect. 3, using ISE, IAE and ITAE as the fitness functions.
minimize the effects of uncertainties in rule-base FLSs To evaluate the influence of the FOU size in the optimi-
(Mendel and John 2002). Some research works are devoted zation search process, in this test we choose three different
to solve real world applications in different areas, for range values for the FOU and in each case, the MFs were
example, in signal processing type-2 fuzzy logic is applied optimized for a 24 db Signal to Noise Ratio (SNR).
in prediction in MackeyGlass chaotic time-series with Finally, in the third set of tests, the MFs of the FLC were
uniform noise presence (Mendel 2000; Karnik and Mendel optimized for a 24 db SNR using the HEM as a global
1999: Castro et al. 2009). In medicine, an expert system optimization method. As in the second set of tests, three
was developed for solving the problem of Umbilical Acid different range values for the FOUs were used, but in this
Base (UAB) assessment (Ozen and Garibaldi 2003). In case, we used the average of two type-1 fuzzy systems to
industry, type-2 fuzzy logic and neural networks are used implement the type-2 FLC.
in the control of non-linear dynamic plants (Melin and This paper is organized as follows: Sect. 2 presents an
Castillo 2002, 2003; Castillo and Melin 2004; Mizumoto introductory description of type-1 and type-2 FLCs and
and Tanaka 1976; Melin and Castillo 2004; Hagras 2004). the performance criteria for evaluating the transient and
In all of these previous works, the type-2 fuzzy systems steady state closed-loop response in a control system.
have been designed manually, with no automatic design Section 3 describes the HEM, which is an intelligent
based on optimizing an objective criterion (Castillo and global optimization method; Sect. 4 is devoted to show
Melin 2007). In this paper, an automatic design procedure the simulation results; in this section, we are showing
for type-2 fuzzy systems is proposed based on the use of an details of the implementation of the feedback control
evolutionary algorithm. system used, we are presenting results from several
This paper deals with the optimization of interval type-2 experiments, the plant was tested using several signal to
membership functions in a fuzzy logic controller (FLC), noise ratio, and we are including a performance com-
the behavior of the FLC after optimization of the MFs parison between type-1 and type-2 fuzzy logic controllers,
under different range values for the Footprint of Uncer- versus optimized type-2 FLCs. An analysis of the results
tainty (FOU) and different noise values is presented. It is a of optimized MFs for different ranges of the FOU and
known fact, that in the control of real systems, the instru- different noise levels is also presented. Section 5 presents
mentation elements (instrumentation amplifier, sensors, a discussion about the results; finally, in Sect. 6, we have
digital to analog, analog to digital converters, etc.) intro- the conclusions.
duce in the collected information some sort of unpredict-
able values (Castillo and Melin 2001). The controllers
designed under idealized conditions tend to behave in an 2 Fuzzy controllers
inappropriate manner (Castillo and Melin 2003). Since,
uncertainty is inherent in real world applications (Sepulv- 2.1 Type-1 fuzzy controllers
eda et al. 2007), we study the effects of uncertainty using a
set of comparative tests for type-1 and type-2 FLCs, in Soft computing techniques have been applied recently in
order to determine which method can offer the most reli- the design of intelligent controllers (Jang et al. 1997).
able control output for a given input. These techniques have tried to avoid the above-mentioned
In the first set of tests, an interval type-2 FLC is used to drawbacks, and they allow us to obtain efficient controllers,
measure the effect of uncertainty and compare it with the which utilize the human experience instead of the con-
results of using a type-1 FLC. We are making the com- ventional mathematical approach (Zadeh 1971, 1973,
parison with experimental results, qualitative observations, 1975a, b). In the cases in which a mathematical represen-
and quantitative measures of errors (Castillo and Melin tation of the controlled systems cannot be obtained, it is

123
Optimization of interval type-2 fuzzy logic controllers using evolutionary algorithms 1147

possible to express the relationships between them, that is,


their process behavior (Sepulveda et al. 2007).
A FLS described completely in terms of type-1 fuzzy
sets is called a type-1 fuzzy logic system (type-1 FLS). It is
composed by a knowledge base that comprises the infor-
mation given by the process operator in form of linguistic
control rules; a fuzzification interface, that has the effect of
transforming crisp data into fuzzy sets; an inference sys-
tem, that uses the fuzzy sets in conjunction with the
knowledge base to make an inference by means of a rea-
soning method; and a defuzzification interface, which
translates the resulting fuzzy set to a real control action
using a defuzzification method (Castillo and Melin 2001).
In this paper, the implementation of the fuzzy controller
in terms of type-2 and type-1 fuzzy sets, has two input
variables: the error e(t), the difference between the refer-
ence signal and the output of the process; and the error Fig. 2 Blurred type-1 membership function
variation De(t),
Z Z
et rt  yt 1 A~ lA~ x; u=x; u Jx  0; 1 4
Det et  et  1 2 x2X u2Jx
RR
In Fig. 1, we show the block diagram that was used as a where denotes the union over all admissible input Rvari-
framework to test the different experiments of this paper ables x and u. For discrete universes of discourse is
P
(Mamdani 1993). replaced by (Mendel 2001). In fact Jx  0; 1 repre-
sents the primary membership of x and lA~x; u is a type-1
2.2 Type-2 fuzzy controllers fuzzy set known as the secondary set.
This uncertainty is represented by a region called foot-
If we have a type-1 membership function as in Fig. 2, and print of uncertainty (FOU). When lA~x; u 1; 8 u 2 Jx 
we are blurring it to the left and to the right then, at a 0; 1 we have an interval type-2 membership function, as
specific value x0 ; the membership function value u0 ;, takes shown in Fig. 3. The uniform shading for the FOU repre-
on different values which may not all be weighted the sents the entire interval type-2 fuzzy set and it can be
same, so we can assign an amplitude distribution to all of described in terms of an upper membership function l A~x
those points. By doing this for all x 2 X; we create a three- and a lower membership function lA~x.
dimensional membership functiona type-2 membership A FLS described using at least one type-2 fuzzy set is
functionthat characterizes a type-2 fuzzy set (Mendel, called a type-2 FLS. Type-1 FLSs are unable to directly
2001). A type-2 fuzzy set A;~ is characterized by:
  
A~ x; u; lA~x; u j8x 2 X; 8u 2 Jx  0; 1 3

in which 0  lA~x; u  1: Another expression for A~ is,

Fig. 1 Experimental framework used for testing the fuzzy controllers Fig. 3 Interval type-2 membership function

123
1148 O. Castillo et al.

handle uncertainties, because they use type-1 fuzzy sets uncertainty, and therefore the behavior of the type-2 FLC
that are certain (Mendel and Mouzouris 1999). On the other can be better than their type-1 counterpart.
hand, type-2 FLSs, are very useful in circumstances where
it is difficult to determine an exact membership value, and 2.3 Performance criteria
there are uncertainties because of the real system measures
(Mendel 2000). For evaluating the transient closed-loop response of a
It is known that type-2 fuzzy sets let us model and computer control system, we can use the same criteria that
minimize the effects of uncertainties in rule-based FLS. normally are used for adjusting constants in proportional
Unfortunately, type-2 fuzzy sets are more difficult to use integral derivative (PID) controllers. These are defined as
and understand than type-1 fuzzy sets; hence, their use is (Deshpande and Ash 1988):
not widespread yet. In a broad sense for type-1 FLSs,
uncertainties can be classified in four groups (Mendel and 1. Integral of Square Error (ISE).
John 2002): Z1
ISE et2 dt 5
1. The meanings of the words that are used in the
antecedents and consequents of rules can be uncertain 0

(words mean different things to different people).


2. Consequents may have histogram of values associated
with them, especially when knowledge is extracted 2. Integral of the Absolute value of the Error (IAE).
from a group of experts who do not all agree. Z1
3. Measurements that activate a type-1 FLS may be noisy IAE jetjdt 6
and therefore uncertain. 0
4. The data used to tune the parameters of a type-1 FLS
may also be noisy.
3. Integral of the Time multiplied by the Absolute value
All of these uncertainties affect directly the optimal
of the Error (ITAE).
settings of the fuzzy set membership functions. Type-1
fuzzy sets are not able to directly model knowledge vari- Z1
ations in crisp membership functions that are totally crisp. ITAE tjetjdt 7
On the other hand, type-2 fuzzy sets are able to model such 0
uncertainties because their membership functions are P R
In our case we will use ; instead of : Each measure
themselves fuzzy.
is based in the error accumulation, and the selection
Similar to a type-1 FLS, a type-2 FLS includes
depends on the type of response desired, errors will
fuzzifier, rule base, fuzzy inference engine, and output
contribute different for each criterion, so we have that large
processor. The output processor includes type-reducer
errors will increase the value of ISE more heavily than to
and defuzzifier; it generates a type-1 fuzzy set output
IAE. ISE will favor responses with smaller overshoot for
(from the type-reducer) or a crisp number (from the
load changes, but ISE will give longer settling time. In
defuzzifier) (Mendel 2005; Karnik and Mendel 2001a, b).
ITAE, time appears as a factor, and therefore ITAE will
A type-2 FLS is again characterized by IFTHEN rules,
penalize heavily errors that occur late in time, but virtually
but its antecedent or consequent sets are now of type-2.
ignore errors that occur early in time. Designing using
Type-2 FLSs, can be used when circumstances are so
ITAE will give us the shortest settling time, but it will
uncertain that it is difficult to determine exact member-
produce the largest overshoot among the three criteria
ship grades such as when training data is corrupted by
considered. Designing considering IAE will give an
noise.
intermediate result, in this case, the settling time will not
In this paper, we are simulating the fact that the
be as large than with ISE and not so small than using ITAE,
instrumentation elements (instrumentation amplifier, sen-
and the same applies for the overshoot response. The
sors, digital to analog, analog to digital converters, etc.) are
selection of a particular criterion depends on the type of
introducing some sort of unpredictable values in the col-
desired response.
lected information (Castillo and Melin 2004). In the case of
the implementation of the type-2 FLC, we have the same
characteristics as in type-1 FLC, but we used type-2 fuzzy 2.4 Main idea of paper
sets as membership functions for the inputs and for the
output. The type-2 fuzzy sets should be designed appro- This paper deals with the optimization of interval type-2
priately so that they can capture the corresponding membership functions in a fuzzy logic controller, the

123
Optimization of interval type-2 fuzzy logic controllers using evolutionary algorithms 1149

behavior of the FLC after optimization of the MFs under found/Pareto Set, VRL Visited Regions List, POS Pareto
different range values for the FOU and different noise Optimal Set.
values is presented. It is a known fact, that in the control of In Fig. 4, we have a general description of HEM con-
real systems, the instrumentation elements (instrumentation taining six main blocks. In the first block, we show that the
amplifier, sensors, digital to analog, analog to digital con- human or group of humans is part of the system. HEM is an
verters, etc.) introduce in the collected information some intelligent evolutionary algorithm that learns from experts
sort of unpredictable values (Castillo and Melin 2001). The their rational and intuitive procedures that they use to solve
controllers designed under idealized conditions tend to optimization problems. In this model, we consider that we
behave in an inappropriate manner (Castillo and Melin have two kinds of humans: real human beings and artificial
2003). Since, uncertainty is inherent in real world appli- humans. In the first block of Fig. 4, we show that real
cations (Sepulveda et al. 2007), we study the effects of human beings form one class. In the second block, the
uncertainty using a set of comparative tests for type-1 and artificial human implemented in the AIIS of the HEM is
type-2 FLCs, in order to determine which method can offer shown. Humans as part of the system are in charge of
the most reliable control output for a given input. teaching the artificial human all the knowledge needed for
A method for designing optimal interval type-2 fuzzy realizing the searching task. HEM has a feedback control
logic controllers using evolutionary algorithms is presented system formed by blocks three and four; and they work
in this paper. Interval type-2 fuzzy controllers can out- coordinately for monitoring and evaluating the evolution of
perform conventional type-1 fuzzy controllers when the the problem to be solved. In the fifth block, we have a
problem has a high degree of uncertainty. However, single objective optimization (SOO) method for solving
designing interval type-2 fuzzy controllers is more difficult single objective optimization problems (SOOP). In addi-
because there are more parameters involved. In this paper, tion, using the SOO method we can to find the ideal, uto-
interval type-2 fuzzy systems are approximated with the pian and nadir vectors for multiple objective optimization
average of two type-1 fuzzy systems, which has been problems (MOOP) (Deb 2002). In the sixth block, we have
shown to give good results in control if the type-1 fuzzy a multiple objective optimization (MOO) method, which is
systems can be obtained appropriately. An evolutionary dedicated to find the Pareto optimal set (POS) in MOOP
algorithm is applied to find the optimal interval type-2 (Kumar and Bauer 2009).
fuzzy system as mentioned above. The human evolutionary
model is applied for optimizing the interval type-2 fuzzy
controller for a particular non-linear plant and results are
compared against an optimal type-1 fuzzy controller.

3 The human evolutionary model

The human evolutionary model (HEM) is a particular type


of an evolutionary algorithm (Montiel et al. 2007). In this
paper, the HEM has been adapted to the problem at hand,
and the details provided here in this paper are for solving
the problem of optimizing fuzzy systems of type-1 and
type-2. The difference between HEM and other evolu-
tionary algorithms is that it includes the human part to
adapt parameters by learning the experience of human
experts on this kind of problems. The HEM model can be
formally defined as follows:
HEM H; AIIS; P; O; S; E; L; TL=PS; VRL; POS 8
where H Human, AIIS Adaptive Intelligent Intuitive Sys-
tem, P Population of size N individuals, O Single or a
multiple objective optimization goals, S Evolutionary
strategy used for reaching the objectives expressed in O, E
Environment, here we can have predators, etc., L Land-
scape, i.e., the scenario where the evolution must be per-
formed, TL/PS Tabu List formed by the bests solutions Fig. 4 General structure of the HEM

123
1150 O. Castillo et al.

A particular characteristic of individuals in HEM is 4 Simulation results


that we are not only including the decision variables;
also, each individual has associated other variables that Figure 1 shows, the feedback control system that was used
are called genetic effects (GE) that will influence the for performing the simulation with the proposed method.
searching process. An individual in HEM is composed of The controller was implemented in the Matlab environment
three parts: (Ingle and Proakis 2000) where it was designed to follow
A genetic representation (gr), that can be codified using the input as close as possible. The non-linear plant was
binary or floating-point representation. Decision variables modeled using Eq. 9
are codified in this part.
A set of genetic effects (ge), that are attributes of each yi 0:2  yi  3  0:07yi  2 0:9  yi  1
individual such as physical structure, gender, actual 0:05  ui  1 0:5  ui  2 9
age, maximum age allowed, pheromone level, etc.
For example, the genetic attribute gender of individual n at The controllers output was directly applied to the
generation x is defined as ge3n;x gegendern;x ; and the plants input. Since we are interested in comparing the
actual age as ge4n;x ; etc. In general, we have the defini- performance between type-1 and type-2 FLC systems, and
tion given by the next expression, where we have omitted optimized interval type-2 FLCs versus optimized type-2
the generation number for simplicity FLCs, using the average of two type-1 FLCs, under
different ranges of the FOU and for 24 db of SNR, we have
gen gemin Strn ; gemax Strn ; gegendern ; the following four cases:
geactAgen ; gemax Agen ; gephLeveln ; . . .; gemn
1. Considering the system as ideal, that is, we did not
introduce any source of uncertainty to the modules of
The third part in the individual representation is devoted
the control system. See experiments 1 and 2.
to the individuals objective values. Objective values are
2. Simulating the effects of uncertain modules (subsys-
codified in vector form, the size of the vector is determined
tems) response introducing some uncertainty at differ-
by the number of objectives that the problem requires. For
ent noise levels. See experiments 3, 4 and 5.
single objective (SO) problems, we will use only one ov
3. Optimizing the MFs of an interval type-2 FLC with
value. For multiple objectives (MO) problems, we will use
different FOU sizes for 24 db of SNR and then obtain
ov(1),,ov(M). Figure 5 shows an individual  pn
ISE, IAE and ITAE for different noise levels. See
grn ; gen ; ovn where grn gr1n ; . . .; gr2n is a vector
experiment 6.
(a row) in the GRNQ matrix. The genetic effects gen are
4. Optimizing the MFs of a type-2 FLC using the average
rows in the matrix GENR :
of two type-1 FLCs also for 24 db of SNR and with
In Montiel et al. (2007), the HEM is described in more
different sizes of the FOU, then repeat last part of
technical detail and it is shown why it is a global optimi-
experiment 6. See experiment 7.
zation method (as other evolutionary algorithms are, such
as the GA). In that paper, the HEM is tested with a set of For case 1, as it is shown in Fig. 1, the systems output
benchmark functions and compared with other algorithms, is connected directly to the summing junction, but in the
such as the GA, to verify that it was faster and more second case, the uncertainty was simulated introducing
accurate. For these reasons it was decided to apply HEM to random noise with a normal distribution (the dashed square
the problem of optimizing interval type-2 fuzzy systems in in Fig. 1). We added noise to the systems output y(i) using
control applications, which is a more challenging and the Matlabs function randn which generates random
important problem to consider. numbers with Gaussian distribution. The signal and the
additive noise in turn, were obtained with expression (10),
the result yi was introduced to the summing junction of
the controller system. For experiments 3 and 4 we used
a = 0.05, and in the set of tests for experiment 5 we varied
the a value to obtain different SNR values.

yi yi a  randn 10
We tested the system using as input, a unit step
sequence, free of noise, ri: For evaluating the systems
response and the comparison between type-1 and type-2
fuzzy controllers, we used the performance criteria: ISE,
Fig. 5 Representing one individual in HEM IAE, and ITAE. In Table 3, we summarized the values

123
Optimization of interval type-2 fuzzy logic controllers using evolutionary algorithms 1151

obtained for each criterion considering 200 units of time. Table 1 Characteristics of the inputs and output of the type-1 FLC
For calculating ITAE, we considered a sampling time Variable Term Center c Standard deviation r
Ts 0:1 s.
For experiments 1, 2, 3, and 4 the reference input r is Input e Negative -10 4.2466
stable and noise free. In experiments 3 and 4, although the Zero 0 4.2466
reference appears to be clean, the feedback at the summing Positive 10 4.2466
junction is noisy since we introduced deliberately noise for Input De Negative -10 4.2466
simulating the overall existing uncertainty in the system; in Zero 0 4.2466
consequence, the controllers inputs e (error), and De Positive 10 4.2466
contains uncertainty data. Output cde NG -10 2.1233
In Experiment 5, we tested the systems, type-1 and type- N -5 2.1233
2 FLC, introducing different values of noise g, that is Z 0 2.1233
modifying the signal to noise ratio SNR (Montiel et al. P 5 2.1233
2007), see Eq. 11, PG 10 2.1233
P 2
jsj Psignal
SNR P 2 11
jgj Pnoise 1 xc 2
lA x e2 r 14
Because many signals have a very wide dynamic range
In terms of the upper and lower membership functions,
[37], SNRs are usually expressed in terms of the
A~x;
we have for l
logarithmic decibel scale, SNR(db), as we can see in
Eq. 12, A~x Nc; r2 ; x
l 15
 
Psignal and for the lower membership function lA~x;
SNRdb 10 log10 12
Pnoise lA~x Nc; r1 ; x 16
 2  2
In Table 4, we show for different values of SNR(db), the 12 xc
r2 12 xc
r2
where N c; r2 ; x  e ;e , and N c; r1 ; x 
behavior of ISE, IAE, ITAE for type-1 and type-2 FLCs. In  2
almost all the cases, the results for type-2 FLC are better 12 xc
r1
e ; (Mendel 2000).
than type-1 FLC.
Hence, in the type-2 FLC, for each input we defined three
In type-1 FLC, we selected Gaussian membership
interval type-2 fuzzy Gaussian MFs: negative, zero, positive
functions (Gaussian MFs) for the inputs and for the output.
in the interval [-10 10], as illustrated in Figs. 6 and 7; for
A Gaussian MF is specified by two parameters {c,r}:
computing the output we have five interval type-2 fuzzy
1 xc 2 Gaussian MFs NG, N, Z, P and PG, with uncertain center
lA x e2 r 13 and fixed standard deviations in the interval [-10 10], as
where c represents the MFs center and r determines the
MFs standard deviation.
For each input of the type-1 FLC, e and De, we defined
three type-1 fuzzy Gaussian MFs: negative, zero, positive.
The universe of discourse for these membership functions
is in the range [-10 10]; their centers are -10, 0 and 10,
respectively, with the same standard deviation of 4.24 for
all of them.
For the output of the type-1 FLC, we have five type-1
fuzzy Gaussian MFs: NG, N, Z, P and PG. These are in the
interval [-10 10], their centers are -10, -0.5, 0, 5, and 10,
respectively; and with the same standard deviation of
2.1233. Table 1 illustrates the characteristics of the inputs
and output of the type-1 FLC.
For the type-2 FLC, as in type-1 FLC we also
selected Gaussian MFs for the inputs and for the output,
but in this case we have interval type-2 Gaussian MFs
with a constant center, c, and an uncertain standard
deviation, r, i.e., Fig. 6 Input e membership functions for the type-2 FLC

123
1152 O. Castillo et al.

of uncertainty to appreciate the effect on the results of the


controllers.
Experiment 1 Ideal system using a Type-1 FLC.
In this experiment, we did not add uncertainty data to
the control system and in Table 3 we show the obtained
values of the ISE, IAE, and ITAE for this experiment.
Experiment 2 Ideal system using a Type-2 FLC.
In this case, we used the same test conditions of
Experiment 1; but now we implemented the controllers
algorithm with interval type-2 fuzzy logic. The corre-
sponding performance criteria are listed in Table 3.
Experiment 3 System with uncertainty using a Type-1
FLC.
In this case, we simulated, using Eq. 23, the effects of
Fig. 7 Input De membership functions for the type-2 FLC
uncertainty introduced to the system by transducers,
amplifiers, and any other element that in real world
applications affects expected values. In Table 3, we can see
the obtained values for the ISE, IAE, and ITAE perfor-
mance criteria for a simulated 24 db signal to noise ratio.
For the case of an ideal system, the results of type-1
should be equal to type-2 when the FOU is zero. However,
when the FOU is not zero, the interval type-2 is not the
same as the type-1 and the results are not equal. Of course,
the difference is small due to the lack of uncertainty in this
case, as can be appreciated in Table 3.
Experiment 4 System with uncertainty using a Type-2
FLC. In this experiment, we introduced uncertainty into the
system, in the same way as in Experiment 3. In this case,
we used an interval type-2 FLC and we improved those
results obtained with a type-1 FLC (Experiment 3), see
Table 3.
Experiment 5 Varying the signal to noise ratio in Type-1
Fig. 8 Output cde membership functions for the type-2 FLC and Type-2 FLCs.
To test the robustness of the type-1 and interval type-2
FLCs, we repeated experiments 3 and 4 providing different
can be seen in Fig. 8. Table 2 shows the characteristics of noise levels, going from 30 db to 8 db of SNR ratio in each
the inputs and output of the interval type-2 FLC. experiment. In Table 4, we summarized the values for ISE,
IAE, and ITAE considering 200 units of time with a Psignal
4.1 Type-1 and Interval Type-2 FLC for ideal system of 22.98 db in all cases. As it can be seen in Table 4, in
and with uncertainty presence of diverse noise levels, in general the behavior of
the interval type-2 FLC is better than the type-1 FLC.
In this case, we consider four experiments performed to the From Table 4, considering two examples, which are the
FLC. First, we consider in experiments 1 and 2 controlling extreme cases; we have for an SNR ratio of 8 db, in the
an ideal system (no uncertainty) with a type-1 FLC and type-1 FLC the following performance values
with Interval Type-2 FLC. Then, we consider in experi- ISE = 321.1, IAE = 198.1, ITAE = 2234.1; and for the
ments 3 and 4 the control of the system with a particular same case, in type-2 FLC, we have ISE = 299.4,
level of uncertainty with a type-1 FLC and with an Interval IAE = 194.1, ITAE = 2023.1. For 30 db of SNR ratio, we
Type-2 FLC. Finally, in experiment 5 we change the level have for the type-1 FLC, ISE = 8.5, IAE = 25.9,

123
Optimization of interval type-2 fuzzy logic controllers using evolutionary algorithms 1153

Table 2 Characteristics of the inputs and output of the type-2 FLC


Variable Term Center c Standard deviation r1 Standard deviation r2

Input e Negative -10 5.2466 3.2466


Zero 0 5.2466 3.2466
Positive 10 5.2466 3.2466
Input De Negative -10 5.2466 3.2466
Zero 0 5.2466 3.2466
Positive 10 5.2466 3.2466
Output cde NG -10 2.6233 1.6233
N -5 2.6233 1.6233
Z 0 2.6233 1.6233
P 5 2.6233 1.6233
PG 10 2.6233 1.6233

Table 3 Comparison of performance criteria for type-1 and type-2 fuzzy logic controllers for 24 DB SNR values obtained after 200 samples
Performance Criteria Type-1 FLC Type-2 FLC
Ideal System System with uncertainty Ideal System System with uncertainty

ISE 7.65 11.9 6.8 10.3


IAE 17.68 36.2 16.4 32.5
ITAE 62.46 289 56.39 264.2

Table 4 Behavior of type-1 and type-2 fuzzy logic controllers after variation of the SNR values obtained for 200 samples
Noise variation Type-1 FLC Type-2 FLC
SNR (db) SNR SumNoise SumNoise (db) ISE IAE ITAE ISE IAE ITAE

8 6.4 187.42 22.72 321.1 198.1 2234.1 299.4 194.1 2023.1


10 10.058 119.2 20.762 178.1 148.4 1599.4 168.7 142.2 1413.5
12 15.868 75.56 18.783 104.7 114.5 1193.8 102.1 108.8 1057.7
14 25.135 47.702 16.785 64.1 90.5 915.5 63.7 84.8 814.6
16 39.883 30.062 14.78 40.9 72.8 710.9 40.6 67.3 637.8
18 63.21 18.967 12.78 27.4 59.6 559.1 26.6 54.2 504.4
20 100.04 11.984 10.78 19.4 49.5 444.2 18.3 44.8 402.9
22 158.54 7.56 8.78 14.7 42 356.9 13.2 37.8 324.6
24 251.3 4.77 6.78 11.9 36.2 289 10.3 32.5 264.2
26 398.2 3.01 4.78 10.1 31.9 236.7 8.5 28.6 217.3
28 631.5 1.89 2.78 9.1 28.5 196.3 7.5 25.5 180.7
30 1,008 1.19 0.78 8.5 25.9 164.9 7 23.3 152.6

ITAE = 164.9, and for the type-2 FLC, ISE = 7, 4.2 Optimization of the fuzzy controllers with HEM
IAE = 23.3, ITAE = 152.6.
These values indicate a better performance of the type-2 In this case, we consider an experiment performed to the
FLC with respect to the type-1 FLC, because the errors are FLC in which now the controllers are optimized with
consistently lower for the interval type-2 fuzzy controller HEM. In Experiment 6, we show the results of optimizing
with different noise levels. the FOU of the membership functions with HEM to

123
1154 O. Castillo et al.

improve the performance of the interval type-2 fuzzy 4.3 Optimization of the interval type-2 fuzzy controller
controller. Of course, the comparison is also with an with the average of two type-1 fuzzy systems using
optimized version of the type-1 fuzzy controller. HEM
Experiment 6 Optimizing the interval type-2 MFs of the
In this case, we use the approximation of an interval type-2
inputs of the FLC for 24 db of SNR, for different ranges of
fuzzy system using the average of two type-1 fuzzy sys-
the FOU.
tems that are found using HEM. The use of the HEM
To evaluate the effects of varying the size of the FOU algorithm is to find the optimal Type-1 FLCs so that the
in the optimization of the type-2 MFs, for 24 db signal average of these type-1 fuzzy systems can better approxi-
to noise ratio, we established different search intervals mate the values of the real interval type-2 FLC. In
for the shadow of the MFs. We maintain the centers Experiment 7, we describe the details of this case for dif-
constant and the upper standard deviation of the ferent FOU values.
Gaussian MFs of the inputs and the lower standard
Experiment 7 Optimizing the interval Type-2 MFs of the
deviations were varied.
FLC for 24 db of SNR, using the average of two Type-1
After using HEM as the optimization method, and tak-
FLCs, varying the FOU.
ing ISE as the fitness function, we found the best values of
the MFs, as can be seen in Table 5. To optimize the interval type-2 MFs of the FLC, we
We started with a narrow interval and finished with simulated the system using two type-1 FLCs. We main-
the wider one. The first interval was in the range of tained constant the centers and upper standard deviations of
3.744.75. After optimization, we calculated the ISE, the Gaussian MFs of the inputs, and we varied the lower
IAE and ITAE values for the noise levels from 8 to 30 values of the standard deviations. After optimization and
db of SNR. taking ISE as the fitness function, we found the best values
The next step was to increase the search interval as of the MFs. For performing the optimization of the con-
follows, an interval between 3.24 and 5.25 for the terms of trollers, we used again the HEM as the optimization
inputs e and De; and after the optimization the ISE, IAE method.
and ITAE values were calculated. For varying the range of the shadow of the FOU we
Finally, the broader search interval was used, between repeated the steps of Experiment 6. Table 7 shows the
2.74 and 5.75 for all the MFs of the inputs terms. In optimized values for the standard deviations of the MFs
Table 5, we can see the optimized values for the standard and in Table 8 we can see the comparison between the
deviations of the MFs and in Table 6 we have the obtained results obtained for the ISE, IAE and ITAE for each vari-
values for the ISE, IAE and ITAE. ation of the FOU.

Table 5 Comparison of the characteristics of the optimized MFs of the type-2 FLC for different intervals of the FOU, for 24 DB of SNR
Variable Type-2 FLC Intervals of the MFs of e, Dea Type-2 FLC Intervals of the MFs of e, Deb Type-2 FLC Intervals of the MFs of e, Dec
Center Standard Standard Center Standard Standard Center Standard Standard
c1 deviation r1 deviation r2 c1 deviation r1 deviation r2 c1 deviation r1 deviation r2

Input e -10 4.75 3.74 -10 5.25 3.2400 -10 5.75 2.9307
0 4.75 4.7131 0 5.25 4.9291 0 5.75 5.1232
10 4.75 3.74 10 5.25 3.2400 10 5.75 3.0478
Input De -10 4.75 4.74 -10 5.25 5.2400 -10 5.75 5.3438
0 4.75 4.6743 0 5.25 4.3890 0 5.75 4.6317
10 4.75 4.7397 10 5.25 5.2400 10 5.75 5.3430
Output cde -10 2.6233 1.6233 -10 2.6233 1.6233 -10 2.6233 1.6233
-5 2.6233 1.6233 -5 2.6233 1.6233 -5 2.6233 1.6233
0 2.6233 1.6233 0 2.6233 1.6233 0 2.6233 1.6233
5 2.6233 1.6233 5 2.6233 1.6233 5 2.6233 1.6233
10 2.6233 1.6233 10 2.6233 1.6233 10 2.6233 1.6233
a
r between 3.74 and 4.75 for the three terms
b
r between 3.24 and 5.25 for the three terms
c
r between 2.74 and 5.75 for the three terms

123
Optimization of interval type-2 fuzzy logic controllers using evolutionary algorithms 1155

Table 6 Performance comparison of the type-2 FLC with optimized MFs for different intervals of the FOU, for 24 DB of SNR
SNR Type-2 FLC Intervals of the MFs of e, Dea Type-2 FLC Intervals of the MFs of e, Deb Type-2 FLC Intervals of the MFs of e, Dec
(db)
ISE IAE ITAE ISE IAE ITAE ISE IAE ITAE

8 108.3 116.6 1131.2 97.6 111.9 1085.3 95.5 110.9 1078.8


10 68.8 93.3 900.6 62.3 89.3 863.9 61.1 88.6 859.9
12 43.9 74.7 715.9 39.9 71.5 687.1 39.2 71 684.8
14 28.2 59.8 568.1 25.7 57.3 545.7 25.3 56.9 544.5
16 18.4 47.9 450.6 16.8 45.9 433.2 16.6 45.7 432.6
18 12.3 38.5 357.7 11.3 37 344.3 11.1 36.8 344
20 8.6 31.2 284.5 7.8 29.9 274 7.7 29.8 273.9
22 6.2 25.5 226.6 5.7 24.3 218.2 5.6 24.3 218.3
24 4.761 21 180.8 4.328 20 174.1 4.3014 20 174.4
26 3.9 17.5 144.7 3.5 16.6 139.2 3.5 16.7 139.7
28 3.3 14.9 116.4 3.0 14 111.6 3 14 112.2
30 3 12.9 94.2 2.7 11.9 89.9 2.7 12 90.4
Values obtained with 200 samples and 60 generations
a
r between 3.74 and 4.75 for the three terms
b
r between 3.24 and 5.25 for the three terms
c
r between 2.74 and 5.75 for the three terms

Table 7 Comparison of the characteristics of the optimized MFs of an average of two type-1 FLCs for different intervals of the FOU, for 24 DB
of SNR
Variable Average of two Type-1 FLCs Intervals Average of two Type-1 FLCs Intervals Average of two Type-1 FLCs Intervals
of the MFs of e, Dea of the MFs of e, Deb of the MFs of e, Dec
Center Standard Standard Center Standard Standard Center Standard Standard
c1 deviation r1 deviation r2 c1 deviation r1 deviation r2 c1 deviation r1 deviation r2

Input e -10 4.75 3.74 -10 5.25 3.2400 -10 5.75 2.9067
0 4.75 4.0771 0 5.25 4.1052 0 5.75 4.8168
10 4.75 3.74 10 5.25 3.3593 10 5.75 3.1084
Input De -10 4.75 4.74 -10 5.25 5.2400 -10 5.75 5.0186
0 4.75 4.436 0 5.25 3.6530 0 5.75 5.0756
10 4.75 4.7398 10 5.25 5.2375 10 5.75 5.0183
Output cde -10 2.6233 1.6233 -10 2.6233 1.6233 -10 2.6233 1.6233
-5 2.6233 1.6233 -5 2.6233 1.6233 -5 2.6233 1.6233
0 2.6233 1.6233 0 2.6233 1.6233 0 2.6233 1.6233
5 2.6233 1.6233 5 2.6233 1.6233 5 2.6233 1.6233
10 2.6233 1.6233 10 2.6233 1.6233 10 2.6233 1.6233
a
r between 3.74 and 4.75 for the three terms
b
r between 3.24 and 5.25 for the three terms
c
r between 2.74 and 5.75 for the three terms

5 Discussion of results different ranges of the FOU. In Figs. 9 and 10, we can see
the optimized MFs that achieved the best results.
5.1 Analysis of the results obtained with the interval For the different noise levels and FOUs, in Table 6, the
type-2 FLC values for the ISE, IAE and ITAE are shown. The best results
were obtained using the wider FOU, i.e. when the search of
The results of Experiment 6 are summarized in Table 5, the optimal values was made in the range 2.745.75.
where we have the optimal values for the MFs parameters In Fig. 13, we can see the ISE errors obtained in each
of the interval type-2 FLC inputs found for the three range of the FOU, for the interval type-2 FLC, with the

123
1156 O. Castillo et al.

Table 8 Performance comparison of the average of two type-1 FLCs with optimized MFs for different intervals of the FOU for 24 DB of SNR
values obtained with 200 samples and 60 generations
SNR (db) Average of two Type-1 FLCs Average of two Type-1 FLCs Average of two Type-1 FLCs Intervals
Intervals of the MFs of e, Dea Intervals of the MFs of e, Deb of the MFs of e, Dec
ISE IAE ITAE ISE IAE ITAE ISE IAE ITAE

8 104.4 114.9 1113.1 95.6 110.8 1074.8 96.1 111 1074.4


10 66.4 91.8 886.3 61.1 88.4 855.3 61.3 88.59 854.6
12 42.4 73.5 704.5 39.1 70.7 680 39.3 70.7 679.3
14 27.3 58.8 558.9 25.2 56.6 539.9 25.3 56.6 539.3
16 17.8 47.1 443 16.5 45.4 428.5 16.5 45.3 427.9
18 11.9 37.8 351.6 11 36.5 340.5 11 36.5 339.9
20 8.2 30.6 279.5 7.6 29.5 270.9 7.6 29.5 270.4
22 6 24.9 222.5 5.5 24 215.6 5.5 23.9 215.1
24 4.5619 20.5 177.4 4.2024 19.7 171.9 4.195 19.6 171.4
26 3.7 17 141.8 3.4 16.3 137.2 3.4 16.2 136.8
28 3.2 14.4 113.8 2.9 13.7 109.9 2.9 13.6 109.5
30 2.9 12.4 92 2.6 11.6 88.4 2.6 11.6 88.1
a
r between 3.74 and 4.75 for the three terms
b
r between 3.24 and 5.25 for the three terms
c
r between 2.74 and 5.75 for the three terms

Fig. 9 Optimized MFs of the input e for the type-2 FLC, for a range
of the FOU between 2.74 and 5.75 Fig. 10 Optimized MFs of the input delta e of the type-2 FLC, for a
range of the FOU between 2.74 and 5.75
optimized parameters of the MFs. The ISE T2-1 corresponds
to the FOU range of 3.744.75, the ISE T-2 to the FOU range
of 3.244.75, and finally the ISE T-3 represents the value
when the range is from 2.74 to 5.75. The difference between In Table 7, we show the optimal values of the MFs of
the ISE T-2 and ISE T-3 is minimum, see Table 6. the two type-1 FLCs for each FOU search range. Fig-
ures 11 and 12 show the optimized MFs that achieved the
5.2 Analysis of the results obtained with the average best results, and as in Experiment 6, they were obtained for
of two Type-1 FLCs the wider search range.
In Table 8, we have the values for the ISE, IAE and
In order to evaluate the performance of an interval type-2 ITAE errors, for the different noise levels and FOU. In this
FLC using the average of two type-1 FLCs, experiment 7 case, as in Experiment 6, the best results were obtained
was done under the same conditions of Experiment 6. with the wider FOU, in the range between 2.74 and 5.75.

123
Optimization of interval type-2 fuzzy logic controllers using evolutionary algorithms 1157

Fig. 11 Optimized MFs of the input e for the average of two type-1
FLCs, for a range of the FOU between 2.74 and 5.75
Fig. 13 Comparison of the ISE errors for the optimized interval type-
2 FLCs. The ISE T2-1 corresponds for a search interval of the FOU of
3.744.75, the ISE T2-2 for a FOU of 3.245.25, and the ISE T2-3 for
a FOU of 3.745.75, for 24 db of SNR

Fig. 12 Optimized MFs of the input delta e for the average of two
type-1 FLCs, for a range of the FOU between 2.74 and 5.75

5.2.1 Comparison of the ISE error obtained for different


FOU ranges, using the average of two Type-1 FLCs,
for each different noise level
Fig. 14 Comparison of the ISE errors for the optimized average of
In Figs. 13 and 14, we can see the ISE errors obtained in each two type-1 FLCs. The ISE PROM-1 corresponds for a search interval
range of the FOU, for the average of two Type-1 FLCs, with of the FOU of 3.744.75, ISE PROM2-2 for a FOU of 3.245.25, ISE
the optimized parameters of the MFs. ISE PROM-1 corre- PROM-3 for a FOU of 3.745.75, for 24 db of SNR
sponds for the range between 3.74 and 4.75 of the FOU, ISE
PROM-2 for the range 3.24 and 4.75 of the FOU, finally ISE
PROM-3 represents the value when the range is 2.74 and different noise levels, we first are going to compare the
5.75. In addition, in this case, the difference between the ISE values of the ISE, IAE and ITAE errors obtained with the
PROM-2 and ISE PROM-3 is minimal. optimized parameters of the MFs of the interval type-2
FLC and the average of the two type-1 FLCs. This com-
5.3 Comparison of the results with the interval type-2 parison is made with the best results obtained with each
FLC and the average of two type-1 FLCs FLC, which, in accordance to Tables 6 and 8, correspond
to those obtained for the wider FOU. The second com-
In order to know which system behaves in a better way in parison is made with the standard deviation values and the
the experiments where we simulated uncertainty through variance obtained in each optimization process to get the

123
1158 O. Castillo et al.

Fig. 15 Comparison of the ISE errors for optimized interval type-2 Fig. 16 Comparison of the IAE errors for optimized interval type-2
FLC and the optimized average of two type-1 FLCs, for different FLC and the optimized average of two type-1 FLCs, for different
noise levels. It is observed a minimal advantage of the optimized noise levels. They behave almost the same, but it is observed a
average of two type-1 FLCs for low noise levels minimal advantage of the optimized average of two type-1 FLCs for
low noise levels

optimal parameters of the MFs for the minimal ISE, IAE


and ITAE errors.

5.3.1 Comparison of the ISE, IAE and ITAE errors


obtained with the interval Type-2 FLC and the
average of two Type-1 FLCs

We can see in Tables 6 and 8 that with the average of two


type-1 FLCs optimized with the wider FOU, it was
obtained a minimum advantage in the values of ISE, IAE
and ITAE errors with respect to the optimized interval
type-2 FLC under the same conditions than the type-1 for
the wider FOU. In Figs. 15, 16, and 17 it is shown that this
advantage is more significant for a low noise level.

5.3.2 Analysis of the results of the interval type-2 FLC


Fig. 17 Comparison of the ITAE errors for optimized interval type-2
and the average of the two Type-1 FLCs FLC and the optimized average of two type-1 FLCs, for different
noise levels, which basically shows no difference
We can see in Table 9, the values obtained in the optimi-
zation process of the optimal parameters for the MFs after
30 tests, of the variance, the Standard deviation, best ISE H0 : l1 l2
value, average ISE obtained with the optimized interval H1 : l1 6l2
type-2 FLC, and average ISE with the average of two
optimized type-1 FLCs. In this case, H0 corresponds to the null hypothesis and
To analyze the statistical behavior of the tests realized H1 the alternative hypothesis. The intention is to show if
with the interval type-2 FLC and the average of two type-1 the average of the two type-1 FLCs could be used properly
FLCs, the hypothesis test using the statistical distribution instead of the interval type-2 FLC. This can be
known as the t student was made, where we have the fol- demonstrated if the null hypothesis is accepted, or on the
lowing hypothesis: contrary if the alternative hypothesis is rejected.

123
Optimization of interval type-2 fuzzy logic controllers using evolutionary algorithms 1159

Table 9 Comparison of the variance, the standard deviation, best ISE value, ISE average, obtained with the optimized interval type-2 FLC and
the optimized average of two type-1 FLCs
Search Type-2 FLC Average of two Type-1 FLCs
Interval
Best ISE value ISE Average Standard deviation Variance Best ISE value ISE Average Standard deviation Variance

3.744.75 4.761 4.9942 0.1649 0.0272 4.5619 4.7701 0.1498 0.0224


3.245.25 4.328 4.5060 0.1460 0.0213 4.2024 4.4009 0.1568 0.0246
2.745.75 4.3014 4.4005 0.1653 0.0273 4.1950 4.3460 0.1424 0.0203

Table 10 Results of the t-student test (ideal systems) it is a better choice to select a type-1 FLC
since it works a little better than a type-2 FLC, and it is
Search interval t0 tpdf_r
easier to implement it. It is also well known that a type-1
3.744.75 5.5096 1.6384 9 10-6 FLC can handle nonlinearities, and uncertainties up to
3.245.25 2.6869 0.0125 some extent.
2.745.75 1.3682 0.1556 In the simulation of real systems, with a higher degree of
uncertainty (for example, due to noise in measurements or
other types of noise), we can conclude that lower overshoot
errors and the best settling times are obtained using an
In Table 10, it is shown that as the search interval of the interval type-2 FLC. The results presented in Table 4 show
FOU increases, the average of two type-1 FLCs has the that the performance of this kind of controllers is better
opportunity to perform as well as the interval type-2 FLC under high noise levels.
and can be considered as a good approximation in this We can conclude that using an interval type-2 FLC in
particular situation. This is concluded because the statisti- real world applications can be potentially a good option
cal values decrease in the t test as the search interval since this type of system is a more suitable choice to
increases, which means that both statistical processes look manage high levels of uncertainty, as we can see in the
almost the same. Of course, this is possible because the results shown in Tables 3 and 4.
optimized type-1 FLCs are used in the approximation of We also discovered that optimizing the membership
the interval type-2 FLC. The advantage of this approxi- functions (MFs) for the inputs of an interval type-2 system
mation is that it is simpler to implement two interval type-1 increases the performance of the system for high noise
FLCs than a complete interval type-2 FLC and also the real levels. In addition, when the search interval for optimizing
time responses are faster, which is needed for real world the MFs is wider, we obtained better results in the per-
applications. It remains to be studied if the approximation formance of the system, as can be seen in the ISE, IAE, and
of an interval type-2 FLC with two type-1 FLCs can be ITAE values of Tables 7 and 8, so these results indicate
extended to more general situations or even to other types that the interval type-2 fuzzy system can handle in a better
of problems. way the uncertainty introduced to the control system.
In comparing the results with other ones reported in We have also shown with statistical evidence that the
the literature, the results presented in Tables 5, 6, 7, 8, 9, performance of the optimized average of two interval type-
and 10 of this paper are better than those presented in 1 FLCs can approximate very closely the real behavior of
(Sepulveda et al. 2007) for the same non-linear plant. The an interval type-2 FLC in this particular situation, and can
main reason for this statement is that now in this paper the be considered as a potential good approximation for real
optimization method has been applied to design in an world control applications because it is simpler to imple-
appropriate fashion the type-2 membership functions for ment and can produce faster real time responses.
the problem. An optimal design of the parameter values in Finally, we can mention that the case study of control-
the type-2 fuzzy sets is very important to achieve perfor- ling a particular non-linear plant appears to be an isolated
mance of the fuzzy controller, which is in this case situation, but we think that it represents well a general class
achieved using the HEM. of non-linear plants and that the results would be similar
for many cases, even for problems of other areas of
application. Of course, in control problems is where it may
6 Conclusions be more crucial to have a quick response for real time
applications and this is when using the proposed method of
We can conclude that using the ISE, IAE, and ITAE as an average of two optimized type-1 fuzzy systems to
performance criteria that in systems without uncertainty approximate an interval type-2 fuzzy system may be more

123
1160 O. Castillo et al.

important. We leave as future work the task of considering Martinez R, Castillo O, Aguilar LT (2009) Optimization of interval
more general cases and or other types of application for this type-2 fuzzy logic controllers for a perturbed autonomous
wheeled mobile robot using Genetic Algorithms. Inf Sci
proposed approximation of interval type-2 fuzzy systems. 179:21582174
Melin P, Castillo O (2002) Intelligent control of non-linear dynamic
plants using type-2 fuzzy logic and neural networks. In:
Proceedings of the Annual Meeting of the North American
References Fuzzy Information Processing Society
Melin P, Castillo O (2003) A new method for adaptive model-based
control of non-linear plants using type-2 fuzzy logic and neural
Castillo O, Melin P (2001) Soft computing for control of non-linear networks intelligent control of non-linear dynamic plants using
dynamical systems. Springer, Heidelberg type-2 fuzzy logic and neural networks. In: Proceedings of the
Castillo O, Melin P (2003) Soft computing and fractal theory for 12th IEEE conference on fuzzy systems
intelligent manufacturing. Springer, Heidelberg Melin P, Castillo O (2004) A new method for adaptive control of non-
Castillo O, Melin P (2004) A new approach for plant monitoring linear plants using type-2 fuzzy logic and neural networks. Int J
using type-2 fuzzy logic and fractal theory. Int J Gen Syst Gen Syst 33:289304
33:305319 Mendel JM (1998) Type-2 fuzzy logic systems: type-reduction. In:
Castillo O, Melin P (2007) Type-2 fuzzy logic: theory and applica- Proceedings of IEEE syst., man, cybern. conf., San Diego
tions. Springer, Heidelberg Mendel JM (1999) Computing with words, when words can mean
Castillo O, Melin P (2008) Intelligent systems with interval type-2 different things to different people. In: Int. ICSC Congress
fuzzy logic. Int J Innovat Comput Inf Control 4:771783 Computat. Intell. Methods Applications, Rochester, New York
Castro JR, Castillo O, Melin P, Rodriguez-Diaz A (2009) A hybrid Mendel JM (2000) Uncertainty, fuzzy logic, and signal processing.
learning algorithm for a class of interval type-2 fuzzy neural Signal Process J 80:913933
networks. Inf Sci 179:21752193 Mendel JM (2001) Uncertain rule-based fuzzy logic systems:
Deb K (2002) Multi-objective optimization using evolutionary introduction and new directions. Prentice Hall, New York
algorithms. Wiley, Great Britain Mendel JM (2005) On a 50% savings in the computation of the
Deshpande PB, Ash RH (1988) Computer process control with centroid of a symmetrical interval type-2 fuzzy set. Inf Sci
advanced control applications. Instrument Society of America, 172:417430
USA Mendel JM, John RI (2002) Type-2 fuzzy sets made simple. IEEE
Hagras HA (2004) Hierarchical type-2 fuzzy logic control architec- Trans Fuzzy Syst 10:117127
ture for autonomous mobile robots. IEEE Trans Fuzzy Syst Mendel JM, Mouzouris GC (1999) Type-2 fuzzy logic systems. IEEE
12:524539 Trans Fuzzy Syst 7:643658
Ingle VK, Proakis JG (2000) Digital signal processing using Mizumoto M, Tanaka K (1976) Some properties of fuzzy sets of type-
MATLAB. Brooks/Cole Publishing Company 2. Inf Control 31:312340
Jang JSR, Sun CT, Mizutani E (1997) Neuro-fuzzy and soft Montiel O, Castillo O, Melin P, Diaz AR, Sepulveda R (2007) Human
computing, a computational approach to learning and machine evolutionary model: a new approach to optimization. Inf Sci
intelligence, Matlab Curriculum Series. Prentice Hall, New 177(10):20752098
Jersey Ozen T, Garibaldi JM (2003) Investigating adaptation in type-2 fuzzy
Karnik NN, Mendel JM (1999) Applications of type-2 fuzzy logic logic systems applied to umbilical acid-base assessment. In:
systems to forecasting of time-series. Inf Sci 120:89111 European symposium on intelligent technologies, hybrid systems
Karnik NN, Mendel JM (2001a) Operations on type-2 fuzzy sets. Int J and their implementation on smart adaptive systems (EUNITE
Fuzzy Sets Syst 122:327348 2003), Oulu, Finland
Karnik NN, Mendel JM (2001b) Centroid of a type-2 fuzzy set. Inf Sepulveda R, Castillo O, Melin P, Rodriguez-Diaz A, Montiel O
Sci 132(14):195220 (2007) Experimental study of intelligent controllers under
Karnik NN, Mendel JM, Liang Q (1999) Type-2 fuzzy logic systems. uncertainty using type-1 and type-2 fuzzy logic. Inf Sci
IEEE Trans Fuzzy Syst 7:643658 177:20232048
Karnik NN, Liang Q, Mendel JM (2001) Type-2 fuzzy logic software. Yager RR (1980) Fuzzy subsets of type II in decisions. J Cybern
Available online at http://sipi.usc.edu/mendel/software/ 10:137159
Klir GJ, Yuan B (1995) Fuzzy sets and fuzzy logic: theory and Zadeh LA (1971) Similarity relations and fuzzy ordering. Inf Sci
applications. Prentice Hall, New York 3:177206
Kumar P, Bauer P (2009) Progressive design methodology for Zadeh LA (1973) Outline of a new approach to the analysis of
complex engineering systems based on multiobjective genetic complex systems and decision processes. IEEE Trans Syst Man
algorithms and linguistic decision making. Soft Comput 13:649 Cybern 3:2844
679 Zadeh LA (1975a) The concept of a linguistic variable and its
Liang Q, Mendel JM (2000) Interval type-2 fuzzy logic systems: application to approximate reasoning. Inf Sci 8:4380
theory and design. IEEE Trans Fuzzy Syst 8:535550 Zadeh LA (1975b) The concept of a linguistic variable and its
Mamdani EH (1993) Twenty years of fuzzy control: experiences application to approximate reasoning, Part 1. Inf Sci 8:199249
gained and lessons learn. In: Marks RJ (ed) Fuzzy Logic
Technology and Applications. IEEE Press, New Jersey

123

You might also like