Scientific Research and Essays Vol. 6(31), pp.

6475-6486, 16 December, 2011

Available online at http://www.academicjournals.org/SRE
DOI: 10.5897/SRE11.1059
ISSN 1992-2248 2011 Academic Journals





Full Length Research Paper

A simplified adaptive neuro-fuzzy inference system
(ANFIS) controller trained by genetic algorithm to
control nonlinear multi-input multi-output systems

Omar F. Lutfy*, Samsul B. Mohd Noor and Mohammad H. Marhaban

Department of Electrical and Electronic Engineering, Faculty of Engineering, University Putra Malaysia, Serdang 43400,
Selangor, Malaysia.

Accepted 24 August, 2011

This paper presents a simplified adaptive neuro-fuzzy inference system (ANFIS) controller to control
nonlinear multi-input multi-output (MIMO) systems. This controller uses only few rules to provide the
control actions, instead of the full combination of all possible rules. Consequently, the proposed
controller possesses several advantages over the conventional ANFIS controller especially the
reduction in execution time, and hence, it is more appropriate for real time control. A real-coded genetic
algorithm (GA) was utilized to optimize the premise and the consequent parameters of the ANFIS
controller, instead of the hybrid learning methods that are widely used in the literature. Accordingly, the
necessity for the teaching signal required by other optimization techniques has been eliminated.
Furthermore, the GA was employed to determine the input and output scaling factors for this controller,
instead of the widely used trial and error method. Two nonlinear MIMO systems were chosen to be
controlled by this controller. In addition, the controller robustness to output disturbances was also
investigated and the results clearly showed the notable accuracy and the generalization ability of this
controller. Moreover, the result of a comparative study with a conventional MIMO ANFIS controller has
indicated the superiority of the simplified MIMO ANFIS controller.

Key words: Neuro-fuzzy systems, ANFIS, genetic algorithms, nonlinear MIMO systems.


INTRODUCTION

The nonlinear universal function approximation property
of fuzzy logic (FL) systems and artificial neural networks
(ANNs) qualifies them to be powerful candidates for
identification and control of nonlinear dynamical systems.
As a result, these techniques have been successfully
applied to solve various kinds of problems in control



*Corresponding author. E-mail: [email protected].

Abbreviations: ANFIS, Adaptive neuro-fuzzy inference system;
MIMO, multi-input multi-output; GA, genetic algorithm; FL, fuzzy
logic; ANNs, artificial neural networks; LSE, least square
estimator; SISO, single-input single-output; PID, proportional
integral derivative; PI, proportional integral; FIS, fuzzy inference
system; MF, membership function; UOD, universe of discourse;
ISE, integral square of errors.
system design resulting in the emergence of intelligent
control systems. However, each of these two intelligent
techniques has its own drawbacks which limit its
usefulness for certain situations and not for others. For
instance, fuzzy logic controllers suffer from some
problems like the selection of appropriate membership
functions, the selection of fuzzy if-then rules, and
furthermore how to tune both of them to achieve the
desired performance. On the other hand, ANNs have
some problems such as their black-box nature, the lack
of knowledge representation power, and the selection of
the proper structure and size to solve a specific problem.
In order to overcome these drawbacks, the integration
of both FL systems and ANNs in a unified system has
received great attention in the literature which resulted in
the appearance of a rapidly emerging field of neuro-fuzzy
systems. One of the most widely used neuro-fuzzy
6476 Sci. Res. Essays



systems is the ANFIS network, which was proposed by
Jang (Jang et al., 1997; Jang, 1993). The hybrid learning
rule, which combines the gradient descent technique and
the least square estimator (LSE), was proposed in (Jang,
1993) to optimize the premise and the consequent
parameters of the ANFIS network, respectively. However,
as a supervised learning method, this hybrid learning rule
requires a teaching signal. This teaching signal can not
be easily provided in some situations. For example, in
control system design, when the objective is to utilize the
ANFIS network as a feedback controller, the teaching
signal represents the desired control actions which are
simply unknown.
In order to alleviate this difficulty, some ANFIS learning
approaches were proposed in the literature. One of these
learning approaches to train the ANFIS as a controller is
represented by the inverse learning method in which the
ANFIS network is trained to acquire the inverse dynamics
of the plant to be controlled. This approach has been
successfully employed to control nonlinear MIMO
systems (Toha and Tokhi, 2009; Yao and Chai, 2007;
Zhou and Jagannathan, 1996). However, the efficiency of
this ANFIS learning approach depends on three crucial
issues. Firstly, the availability of an accurate model for
the process to be controlled, which can not be easily
developed for complex systems; Secondly, the existence
of the inverse dynamics of the system, which is not
always available; and thirdly, the way the training data
are distributed. More precisely, it is desired to make the
training data uniformly distributed across the controller
input space. However, this data distribution may not be
achieved for all the systems due to either the small
number of data or the constraints imposed by the system
dynamics.
Another learning method to train the ANFIS as a MIMO
controller was proposed in (Djukanovi et al., 1997). The
basic idea of this method, which was called temporal
back propagation, is to consider both the controller and
the plant as a single unit at each time step. Several
ANFIS controllers were adopted to handle different
operating points and a neural network classifier was used
to settle on the best controller for a particular operating
point. However, this ANFIS training method is
characterized by its heavy computational load and the
complexity in design and implementation.
In addition to the above mentioned methods, a different
training technique in training the ANFIS as a controller for
nonlinear MIMO systems was utilized in (Mahmoud et al.,
2010; Cao et al., 2007; Touati et al., 2002). This
technique is based on training the ANFIS network by
using data which are collected from another working
controller. The disadvantage of this method is that, the
effectiveness of the resulting ANFIS controller depends
heavily on the performance of the original controller,
which is mostly a linear one.
One of the most powerful approaches to train a neuro-
fuzzy system to act as a feedback controller is




represented by using derivative-free methods as the
optimization technique to train the controller. These
derivative-free methods do not need the teaching signal
information in training the neuro-fuzzy system as a
controller. The genetic algorithm is one of the most widely
used derivative-free techniques and many researchers
have successfully utilized it in various control applications
to solve the teaching signal problem.
For example, some researchers used binary-coded GA
to tune a neuro-fuzzy controller utilizing off-line learning
algorithms (Chou, 2006; Seng et al., 1999). In addition,
they used the grid partitioning method to define the
structure of the neuro-fuzzy controller. However, the grid
partitioning method leads to the exponential increase in
the number of fuzzy rules as the number of input
variables increases. Moreover, the binary-coded GA
requires encoding and decoding operations to construct
the chromosomes of the GA which result in some
resolution problems, the increment in chromosome
lengths, and the need to select the values of some user-
defined upper and lower limits of the genes.
Zhou and Lai (2000) used real-coded GA to train a
fuzzy controller to control a linear single-input single-
output (SISO) system. They also used the grid
partitioning method to define the structure of the fuzzy
controller. Their control system consisted of a
combination of a classical proportional-integral-derivative
(PID) controller as the master controller and the fuzzy
controller as the slave controller to enhance the
performance of the PID controller.
Serra and Bottura (2006) proposed a gain scheduling
adaptive control scheme based on a genetically-tuned
fuzzy proportional-Integral (PI) controller.
It is worth highlighting that, in (Chou, 2006; Serra and
Bottura, 2006; Zhou and Lai, 2000; Seng et al., 1999) the
controller was used to control SISO systems only. Hence,
the applicability of these control algorithms to control
MIMO systems is questionable.
In order to handle the teaching signal problem in this
work, a real-coded GA, which does not use a teaching
signal, is utilized to optimize the premise and the
consequent parameters of a simplified ANFIS structure,
which was proposed in (Lutfy et al., 2011; Lutfy et al.,
2010) to control (SISO) systems. In the present work, this
controller is trained to act as a PID-like feedback
controller for nonlinear MIMO systems. This simplified
ANFIS controller requires less training time and memory
resources compared to the conventional ANFIS
controller. Moreover, the input and output scaling factors
for this simplified ANFIS controller are also determined by
the real-coded GA.


Structure of the simplified pid-like adaptive neuro-
fuzzy inference system (ANFIS) controller

In a conventional fuzzy reasoning system, if there are n
Exhau
st
openi
ng

Lutfy et al. 6477





Figure 1. Structure of the simplified PID-like ANFIS controller.



input variables each of which is partitioned into p fuzzy
sets, then p
n
fuzzy rules are required to construct the
corresponding fuzzy inference system (FIS) for a
particular application. As a fuzzy reasoning system, the
ANFIS network utilizes the Sugeno fuzzy models, which
can be either zero-order or first-order models depending
on the consequent part of the fuzzy rules. For both of
these models, the total number of fitting parameters in
the ANFIS structure is given by:

f(n,p,m)=n.p.m+p
n
.(n+1) for the first-order Sugeno fuzzy
model (1)

f(n,p,m)=n.p.m+p
n
for the zero-order Sugeno fuzzy
model (2)

Where m is the number of fitting parameters per
membership function (MF) (Jovanovi et al., 2004). In this
work, the zero-order Sugeno fuzzy model is used since it
requires less number of parameters compared to the first-
order Sugeno fuzzy model. Therefore, according to
Equation 2, if three input variables, seven fuzzy sets per
input, and two fitting parameters per MF are used, then
the resulting number of parameters will be 385.
Obviously, this number of parameters will complicate the
task of the optimization technique in finding the optimal
values for these parameters in terms of accuracy and
computational complexity, not to mention the case of
MIMO systems which requires more ANFIS parameters.
The so called curse of dimensionality can be evidently
seen from the p
n
term included in Equations 1 and 2. In
order to overcome this dimensionality problem, the
ANFIS structure (Lutfy et al., 2011; Lutfy et al., 2010)
shown in Figure 1 is used in this work.
In this structure, the input variables
1
x ,
2
x , and
3
x are
the error of the plant being controlled, the error rate of
change, and the summation of errors respectively, while
the output variable, y, is the control action of this ANFIS
controller. Seven MFs are used for each of the three
input variables to the ANFIS controller. This number of
MFs was found to be adequate to achieve satisfactory
control results, where increasing this number into eleven
terms did not improve the performance significantly. On
the other hand, when five terms for each input variable
were used the control accuracy has deteriorated. In the
following, to describe the function of each layer in the
ANFIS structure, the output of the ith node in layer k is
expressed as O
k,i
.

Layer 1: Each node in this layer is used to assign the
degree of membership for each input variable to the
ANFIS network. In particular, the function of each node is
defined by the following:

21 ..., , 16 , 15 ), (
14 ..., , 9 , 8 ), (
7 ..., , 2 , 1 ), (
3 , 1
2 , 1
1 , 1
14
7
= =
= =
= =

i x O
i x O
i x O
i
i
i
A i
A i
A i

(3)

The bell-shaped activation functions are used in this work
to be the MFs for each input variable. These bell-shaped
6478 Sci. Res. Essays



activation functions are defined by the following
expression:

|
|
.
|

\
|
=
2
) ( 2 / 1 exp ) (
k
a
k
a
a
x
A
x
A k
k A
C x
x
o
(4)

where x
k
, k = 1, 2, 3, are the scaled input variables after
multiplying them by the input scaling factors (c
1
, c
11
, and
c
111
) as illustrated in Figure 1.
k
a
x
A
C and
k
a
x
A
o are the
centers and the widths of the MFs, respectively, and they
are known as the premise parameters.

Layer 2: Each node in this layer generates the firing
strength of the corresponding fuzzy rule utilizing the
multiplication operation. As can be seen from Figure 1,
this simplified ANFIS structure uses only seven fuzzy
rules in this layer, instead of the 343 rules used in the
conventional ANFIS structure (that is, p
n
, where p = 7
and n = 3), see Equation 2. Consequently, the output of
Layer 2 is expressed by:

7 ,..., 2 , 1
14 , 1 7 , 1 , 1 , 2
= =
+ +
i O O O O
i i i i
(5)

Layer 3: This layer has seven nodes as well. The ith node
in this layer determines the ratio of the ith rules firing
strength to the sum of the firing strengths of all the rules,
as expressed by the following:

=
= =
7
1
, 3
j
j
i
i i
w
w
w O where i=1 to 7. (6)

Layer 4: Similar to Layers 2 and 3, this layer has also
seven nodes. Each node in this layer multiplies the
corresponding output from Layer 3 by the consequent
parameter of the ANFIS structure. The output of each
node is given by:

i i i
k w O
0 , 4
=
(7)

where i=1 to 7,
i
w is the ith output from Layer 3, and k
0i
is
the ith consequent parameter of the ANFIS structure.

Layer 5: This layer includes only one node which
calculates the overall output as follows:

=
=
7
1
, 4 5
i
i
O O (8)

Finally, O
5
is multiplied by the output scaling factor (c
2
) to
determine the final output of this ANFIS controller, as




expressed by the following:

5 2
O c u y = = (9)

For the SISO case, 53 parameters are required to
represent this simplified ANFIS structure in the GA (49
modifiable parameters plus 4 scaling factors). In this
work, two ANFIS controllers are used to control the
MIMO plants. Each of these two controllers has exactly
the same structure described above. As a result, 106
parameters are used to represent the simplified MIMO
ANFIS controller in the GA.


The real-coded genetic algorithm (GA)
implementation to train the simplified multi-input
multi-output (MIMO) adaptive neuro-fuzzy inference
system (ANFIS) controller

The MATLAB software was used in this work to
implement the real-coded GA for training the simplified
MIMO ANFIS controller using a specific program which
has been written in an m-file. As mentioned before, each
chromosome in the GA consists of 53 genes which
represent all the ANFIS parameters. These genes
include 3 genes for the input scaling factors, 1 gene for
the output scaling factor, 42 genes for the premise
parameters, and 7 genes for the consequent parameters
of the ANFIS structure. The universe of discourse (UOD)
for each input variable was selected to be from -6 to 6,
keeping in mind that other range can also be used since
there are input and output scaling factors. Four real-
coded GA operators were used in this work. These
operators include the hybrid selection, the elitism, the
crossover, and the mutation operators. The hybrid
selection operator, which was proposed in (Al-Said,
2000), is a combination of the roulette wheel and the
deterministic selection. In this operator, only the
chromosomes that have better fitness values than the
worst individual in the old population are accepted in the
new population. In the elitism operator, the best two
chromosomes in a given generation are copied directly
into the next generation as they are in order to prevent
the best fitness value in a given generation from
becoming worse than that in the previous one
(Sivanandam and Deepa, 2008). In the crossover
operator, a pair of selected chromosomes exchanges
their information by exchanging a subset of their
components, where an integer position k is selected
uniformly at random along the chromosome length. Then
two new chromosomes are generated by swapping all the
genes from positions k+1 to L, where L is the
chromosome length (Sivanandam and Deepa, 2008).
Finally, the real-coded mutation operator causes random
changes in the components of the new population of
chromosomes by replacing the mutated gene with
another random number chosen in the same range




assigned for that gene.
The procedure of the real-coded GA adopted in this
work to train the simplified MIMO ANFIS controller is
summarized in the following steps:

Step 1: Initialize the crossover probability (P
c
), the
mutation probability (P
m
), the population size, and the
maximum number of generations.

Step 2: Generate randomly the initial population of
chromosomes within certain limits. Each of these
chromosomes represents the entire premise and
consequent parameters along with the input and output
scaling factors for one MIMO ANFIS controller.

Step 3: Evaluate the objective function for each
chromosome in the population using the integral square
of errors (ISE) criterion, which has the following form:

) ( ) ( 5 . 0 5 . 0
2
2
0
2
1
k e k e ISE
T
k
+ =

=
(10)

where e
1
(k) and e
2
(k) are the error between the desired
output and the plant output at sample k for each of the
two outputs of the system (that is, e
1
(k) = r
1
(k) y
1
(k),
e
2
(k) = r
2
(k) y
2
(k)), and T is the observation time. For
the current application, and after a fair amount of
simulation tests, the 0.5ISE was found to be the best
performance index in training the simplified MIMO ANFIS
controller to control nonlinear MIMO systems in this work.
Then, for each chromosome in the current population,
determine the fitness function using the Darwinian fitness
equation which has the following form:

function objective
fitness
+
=
c
1
(11)

wherec is a small constant used to avoid division by
zero.

Step 4: Put in descending order all the chromosomes in
the current population (that is, the first one is the fittest).
Then apply the elitism strategy described before.

Step 5: Select two individuals from the current population
utilizing the hybrid selection method, and then apply the
real-coded crossover and mutation operators described
previously to generate two new chromosomes.

Step 6: Put the resulting two chromosomes in the new
population.

Step 7: Repeat Step 5 until all the chromosomes in the
new population are generated, that is, until the new
population size is equivalent to the initial (old) population
size.
Lutfy et al. 6479



Step 8: Replace the initial (old) population with the new
population.

Step 9: Stop if the maximum number of generations is
reached, and the first chromosome in the last generation
is the optimal controller found by the GA, otherwise
increase the generation counter by one and go to step 3.


RESULTS AND DISCUSSION

In order to evaluate the effectiveness of the proposed
simplified MIMO ANFIS controller, two nonlinear MIMO
plants have been chosen to be controlled by this
controller. Unlike the control of SISO systems in (Lutfy et
al., 2011; Lutfy et al., 2010), the simplified MIMO ANFIS
controller has to deal properly with the variable
interactions in the MIMO systems considered in this work.
In this regard, the objective of the following simulation
tests is to investigate the controller ability in decoupling of
loop-interactions while at the same time tracking the
given testing signals. Therefore, the training and the
testing signals were deliberately selected as shown in
Figures 2 (a, b) and 3 (a, b), respectively. From these
figures, it can be seen that there is a difference between
the training and the testing signals. This difference is
important in order to evaluate the generalization ability of
the simplified MIMO ANFIS controller. The parameters of
the real-coded GA were set to the following values;
population size: 50, maximum number of generations:
300, P
c
: 0.8, and P
m
: 0.05. For the current application,
and after several simulation tests, these parameter
settings were found to be adequate to guarantee a
satisfactory performance in training the simplified MIMO
ANFIS controller by the GA to control the following
nonlinear MIMO plants:

Plant 1: A nonlinear MIMO plant is described by the
following (Narendra and Parthasarathy, 1990):

(

+
(
(
(
(

+

+

=
(

) 1 (
) 1 (
) 1 ( 1
) 1 ( ) 1 (
) 1 ( 1
) 1 (
) (
) (
2
1
2
2
2 1
2
2
1
2
1
k u
k u
k y
k y k y
k y
k y
k y
k y
(12)

Figure 3 shows the output responses, control actions and
the best 0.5ISE against the generations for each output
of this plant.
As can be seen from Figures 3a and b, the simplified
MIMO ANFIS controller has done well in both tracking
and de-coupling in controlling this nonlinear MIMO plant.
More specifically, the tracking of the testing signals was
achieved with zero steady-state error and with some
overshoots at the beginning of each step change. Figure
3c shows a sharp control signal when the testing signal
has changed its amplitude. This sharp control action is
6480 Sci. Res. Essays



r
1
(
k
)

(a)
(b)
r
2
(
k
)













0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0 20 40 60 80 100 120 140 160 180 200
Time samples (k)

-0.9
-0.8
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
0 20 40 60 80 100 120 140 160 180 200
Time samples (k)


Figure 2. Training signals: a, for the first input; b, for the second Input.



necessary to compensate for the interactions between
the two channels of this nonlinear MIMO system, and
consequently, to achieve the good tracking performance
in Figure 3a. On the other hand, Figure 3d shows that the
transition in the second control signal was not as large as
that of the first control signal. Bearing in mind the
difference between the training and the testing signals, it
can be concluded that this controller has achieved good
generalization ability. From Figure 3e, it can be seen that
the 0.5ISE has reached its near optimal value from the
first few generations. Therefore, the selection of the
maximum number of generations to be 300 seems to be
adequate. This fact indicates the fast convergence to the
optimal solution that has been achieved by the GA in
training the simplified MIMO ANFIS controller.

Plant 2: This nonlinear MIMO plant is described by the
following (Petlenkov, 2007):

) 2 ( 2 . 0 ) 1 ( ) 2 ( 5 . 0
) 2 ( ) 1 ( 1
)) 2 ( sin( ) 1 ( 5 . 0
) (
) 2 ( 2 . 0 ) 1 ( ) 2 ( 3 . 0
) 2 ( ) 1 ( 1
) 2 ( ) 1 ( 7 . 0
) (
1 2 2 2
1
2
2
2 2
2
2 1 1 2
2
2
1
1 1
1
+ + +
+ +

=
+ + +
+ +

=
k u k u k u
k y k y
k y k y
k y
k u k u k u
k y k y
k y k y
k y
(13)
Figure 4 shows the simulation results of controlling this
nonlinear MIMO plant.
In spite of the complexity of this MIMO plant, the
performance of the simplified MIMO ANFIS controller was
good in de-coupling of loop interactions and in tracking
the desired testing signals with zero steady-state error
and some oscillations at the start of each change in the
testing signals (Figures 4a and b). Even though the
training signals are different from the testing signals, the
simplified MIMO ANFIS controller has successfully
generalized its learning to effectively handle the
unexpected testing signals. This performance for the
controller is due to its control signal given in Figures 4c
and d. Similar to the first plant, the 0.5ISE has reached to
its near optimal value from the first few generations, as
can be seen from Figure 4e.


Robustness test

In order to assess the robustness of the simplified MIMO
ANFIS controller, a disturbance rejection test was
conducted on Plant 2 using the training and the testing
signals of the previous tests. This robustness test was
done by applying a bounded external disturbance of 20%
Lutfy et al. 6481



(a)
(b)
(c)
r
1
(
k
)
,

y
1
(
k
+
1
)

r
2
(
k
)
,

y
2
(
k
+
1
)

u
1
(
k
)

u
2
(
k
)




















0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0 20 40 60 80 100 120 140 160 180 200

-0.45
-0.4
-0.35
-0.3
-0.25
-0.2
-0.15
-0.1
-0.05
0
0 20 40 60 80 100 120 140 160 180 200

-0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
0 20 40 60 80 100 120 140 160 180 200

-0.4
-0.35
-0.3
-0.25
-0.2
-0.15
-0.1
-0.05
0
0 20 40 60 80 100 120 140 160 180 200

Reference input r
1
(k)
Plant output y
1
(k+1)
Reference input r
2
(k)
Plant output y
2
(k+1)
Time samples (k)
Time samples (k)
Time samples (k)
Time samples (k)

(d)
6482 Sci. Res. Essays



Figure 3. Plant 1 (a) first testing input and plant output (b) second testing
input and plant output (c) first control signal (d) second control
signal (e) best 0.5ISE.
(d)
(e)
B
e
s
t

o
b
j
e
c
t
i
v
e












0
5
10
15
20
25
30
35
40
45
50
0 50 100 150 200 250 300

Generations


Figure 3. Plant 1: a, first testing input and plant output; b, second test ing input and plant
output; c, first control signal; d, second control signal; e, best 0.5ISE.



r
1
(
k
)
,

y
1
(
k
+
1
)

r
2
(
k
)
,

y
2
(
k
+
1
)

u
1
(
k
)

u
2
(
k
)

(a)
(b)
(c)
Figure 4 shows the simulation results of controlling this nonlinear MIMO plant.



















0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0 20 40 60 80 100 120 140 160 180 200

-0.45
-0.4
-0.35
-0.3
-0.25
-0.2
-0.15
-0.1
-0.05
0
0 20 40 60 80 100 120 140 160 180 200

0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0 20 40 60 80 100 120 140 160 180 200

-0.4
-0.35
-0.3
-0.25
-0.2
-0.15
-0.1
-0.05
0
0 20 40 60 80 100 120 140 160 180 200

Time samples (k)
Time samples (k)
Time samples (k)
Time samples (k)
Reference input r
1
(k)
Plant output y
1
(k+1)
Reference input r
2
(k)
Plant output y
2
(k+1)

Lutfy et al. 6483



u
1
(
k
)

u
2
(
k
)

B
e
s
t

o
b
j
e
c
t
i
v
e


(c)
(d)
(e)


















0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0 20 40 60 80 100 120 140 160 180 200

-0.4
-0.35
-0.3
-0.25
-0.2
-0.15
-0.1
-0.05
0
0 20 40 60 80 100 120 140 160 180 200

0
10
20
30
40
50
60
70
0 50 100 150 200 250 300

Time samples (k)
Time samples (k)
Generations


Figure 4. Plant 2 (a) first testing input and plant output (b) second testing (c) first control
signal (d) second control signal (e) best 0.5ISE.



of the first output and -20% of the second output during
only the testing phase, which means that the controller
was not trained to handle these disturbances during the
training phase. The first disturbance was applied for 20
samples at the interval (
49 30 s s k
) while the second
one was applied for 20 samples at the interval
(
149 130 s s k
) of the two testing signals. Figure 5
shows the two plant outputs together with the two control
actions of the simplified MIMO ANFIS controller for this
test.
Referring to Figures 5a and b, it can be seen that the
simplified MIMO ANFIS controller has overcome the
external disturbances, during their effect and after their
disappearance as well, during the two disturbance
periods. Figures 5c and d illustrate the adaptations in
control actions achieved by the ANFIS controller to
6484 Sci. Res. Essays



(a)
(b)
(c)
(d)
r
1
(
k
)
,

y
1
(
k
+
1
)

r
2
(
k
)
,

y
2
(
k
+
1
)

u
1
(
k
)

u
2
(
k
)




















0
0.1
0.2
0.3
0.4
0.5
0.6
0 20 40 60 80 100 120 140 160 180 200

-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
0 20 40 60 80 100 120 140 160 180 200

0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0 20 40 60 80 100 120 140 160 180 200

-0.5
-0.45
-0.4
-0.35
-0.3
-0.25
-0.2
-0.15
-0.1
-0.05
0
0 20 40 60 80 100 120 140 160 180 200

Time samples (k)
Time samples (k)
Time samples (k)
Time samples (k)
Reference input r
1
(k)
Plant output y
1
(k+1)
Reference input r
2
(k)
Plant output y
2
(k+1)


Figure 5. Plant 2 subjected to 20% external disturbances; a, first testing input and plant output;
b, second testing input and plant output; c, first control signal; d, second control signal.
Lutfy et al. 6485



Table 1. Comparison results of the conventional and the simplified MIMO ANFIS controllers

Controll
ed plant
Conventional MIMO ANFIS
controller
Simplified MIMO ANFIS
controller
Average Training
0.5ISE
Average testing
0.5ISE
Average time
(sec.)
Average Training
0.5ISE
Average testing
0.5ISE
Average time
(sec.)
Plant 1 0.422 0.345 652.526 0.200 0.108 89.858
Plant 2 0.184 0.102 649.471 0.168 0.083 90.125



handle these disturbances. This result indicates the
ability of the simplified MIMO ANFIS controller to deal
appropriately with external disturbances.


A comparative study with the conventional multi-
input multi-output (MIMO) adaptive neuro-fuzzy
inference system (ANFIS) controller

As previously mentioned, only few rules, in particular p
rules, were utilized in the rule base of the simplified
ANFIS controller to provide the control actions, instead of
the full combination of all possible rules, that is, the p
n

rules, resulted from applying the grid partitioning method
to partition the input space in the conventional ANFIS
structure, where n is the number of input variables and p
is the number of fuzzy sets for each input variable. As a
result, the simplified ANFIS controller has several
advantages over the conventional ANFIS controller,
particularly the reduction in processing time without
sacrificing the controller performance, as will be seen in
this section. In this regard, it is important to compare the
performances of the simplified and the conventional
MIMO ANFIS controllers, and this section is devoted for
this purpose. The structure of the simplified ANFIS
controller was described at the beginning of this paper.
This structure requires 53 genes in the GA to optimize
the ANFIS parameters to control SISO systems, and 106
parameters to control MIMO systems in this work. The
conventional MIMO ANFIS controller has exactly the
same structure as that of the simplified MIMO ANFIS
controller except that the former requires 343 fuzzy rules
(that is, p
n
rules, where p = 7 and n = 3). Therefore, each
of Layers 2, 3, and 4 of the conventional ANFIS controller
consist of 343 nodes. According to Equation 2, 385
parameters are required to optimize the parameters of
the conventional ANFIS controller to control SISO
systems. On the other hand, for MIMO systems, two of
the above conventional ANFIS controllers are required. In
order to achieve a fair comparison, all the tests in this
study were done under identical conditions for the two
controllers. Specifically, the same training and testing
signals considered before were used for all the
comparison tests. In addition, the real-coded GA, with
exactly the same operators and settings, was used to
train the two controllers. Plants 1 and 2 described before
were used to conduct this comparative study. Due to the
stochastic nature of the GA, the result obtained from one
simulation run might be different from that of other runs.
For this reason, and in order to achieve a more reliable
comparison study, five runs were conducted for each
controller to control each plant and the average objective
function and average time for each of these five runs
were considered as the basis in this comparison. Table 1
above summarizes the results of comparing the simplified
MIMO ANFIS controller with the conventional MIMO
ANFIS controller to control the two plants.
The main advantage of the simplified MIMO ANFIS
controller, which is the reduction in training time, can be
easily seen from Table 1 for the two plants. At the same
time, by carefully examining Table 1, it can be concluded
that the simplified MIMO ANFIS controller not only
outperforms the conventional MIMO ANFIS controller in
terms of training time but also in terms of control
accuracy. More specifically, the average training and
testing objective functions are less for the two plants in
the case of the simplified MIMO ANFIS controller. In fact,
the average testing 0.5ISE indicates the generalization
ability of a given controller, and as can be seen from
Table 1, this average testing 0.5ISE was 0.345 and 0.102
for the conventional ANFIS controller against only 0.108
and 0.083 for the simplified ANFIS controller in controlling
Plants 1 and 2, respectively. This result confirms the fact
that, as the controller complexity increases, its
generalization ability deteriorates, as was highlighted in
(Mascioli and Martinelli, 1998).
It is worth mentioning that although the control
algorithm utilized in this work is based on an off-line
design technique, the training time of the simplified MIMO
ANFIS controller proposed in this work is dramatically
less than that of the conventional MIMO ANFIS controller.
This short processing time is a very desirable attribute in
designing an on-line training method for the controller.
Therefore, an on-line training algorithm, utilizing the same
real-coded GA used in this work, is currently being
developed to provide an on-line self-tuning method for
the simplified MIMO ANFIS controller, and hence to
promote the off-line design method presented in this work
into an adaptive learning approach.


Conclusions

A genetically-trained simplified PID-like ANFIS controller
6486 Sci. Res. Essays



was utilized in this work to control nonlinear MIMO
systems. The GA does not require a teaching signal in its
operations and hence, it is a more suitable technique to
train the ANFIS network as a feedback controller in this
work compared to the commonly used method to train the
ANFIS namely, the hybrid learning method. Moreover,
the real-coded GA was used to determine the optimal
settings for the controller scaling factors, instead of the
widely used trial and error method. The simulation results
showed the effectiveness of the real-coded GA in training
the simplified MIMO ANFIS controller in terms of control
accuracy and generalization ability. Moreover, from the
robustness test, this controller has shown a notable
ability in eliminating the effects of external disturbances
during their effect and after their disappearance as well.
In addition, from a comparative study with a conventional
MIMO ANFIS controller, the simplified MIMO ANFIS
controller has shown its superiority in terms of the
reduced training time and the control accuracy.


REFERENCES

Al-Said IAM (2000). Genetic algorithms based intelligent control. PhD
dissertation, University of Technology, Baghdad, Iraq.
Cao H, Si G, Zhang Y, Ma X (2007). A hybrid controller of self-
optimizing algorithm and ANFIS for ball mill pulverizing system.
Proceedings of the IEEE international conference on mechatronics
and automation, Harbin, China, pp. 3289- 3294.
Chou CH (2006). Genetic algorithm-based optimal fuzzy controller
design in the linguistic space. IEEE Trans. Fuzzy Syst., 14(3): 372-
385.
Djukanovi MB, alovi MS, Veovi BV, obaji DJ (1997). Neuro-
fuzzy controller of low head hydropower plants using adaptive-
network based fuzzy inference system. IEEE T. Ener. Conver, 12(4):
375-381.
Jang JSR (1993). ANFIS: adaptive-network-based fuzzy inference
system. IEEE T. Syst. Man. Cy., 23(3): 665-685.
Jang JSR, Sun CT, Mizutani E (1997). Neuro-fuzzy and soft computing:
A computational approach to learning and machine intelligence.
Prentice-Hall, Englewood Cliffs, New Jersey.
Jovanovi BB, Reljin IS, Reljin BD (2004). Modified ANFIS architecture -
improving efficiency of ANFIS technique. In the 7th seminar on neural
network applications in electrical engineering, Serbia and
Montenegro, pp. 215-220.
Lutfy OF, Mohd Noor SB, Marhaban MH, Abbas KA (2010). A simplified
PID-like ANFIS controller trained by genetic algorithm to control
nonlinear systems. Aus. J. Basic Appl. Sci., 4(12): 6331-6345.






Lutfy OF, Mohd Noor SB, Marhaban MH, Abbas KA (2011). Nonlinear
modeling and control of a conveyor-belt grain dryer utilizing neuro-
fuzzy systems. Proc. Inst. Mech. Eng. Part I-J. Syst. Control Eng.,
225: 611-622.
Mahmoud TS, Marhaban MH, Hong TS (2010). ANFIS: Self-tuning
fuzzy PD controller for twin rotor MIMO system. IEE J. T. Elect.
Elect., 5(3): 369-371.
Mascioli FMF, Martinelli G (1998). A constructive approach to neuro-
fuzzy networks. Signal Process, 64(3): 347-358.
Narendra KS, Parthasarathy K (1990). Identification and control of
dynamical systems using neural networks. IEEE T. Neural Network,
1(1): 4-27.
Petlenkov E (2007). NN-ANARX structure based dynamic output
feedback linearization for control of nonlinear MIMO systems. In
Mediterranean conference on control and automation, Athens-
Greece, pp. 1-6.
Seng TL, Khalid MB, Yusof R (1999). Tuning of a neuro-fuzzy controller
by genetic algorithm. IEEE T. Syst. Man Cy. B., 29(2): 226-236.
Serra GLO, Bottura CP (2006). Multiobjective evolution based fuzzy PI
controller design for nonlinear systems. Eng. Appl. Artif. Intell., 19(2):
157-167.
Sivanandam SN, Deepa SN (2008). Introduction to genetic algorithms.
Springer-Verlag, Berlin, Heidelberg.
Toha SF, Tokhi MO (2009). Dynamic nonlinear inverse-model based
control of a twin rotor system using adaptive neuro-fuzzy inference
system. Third UKSim European symposium on computer modeling
and simulation, Athens, Greece, pp. 107-111.
Touati Y, Djouani K, Amirat Y (2002). Neuro-fuzzy based approach for
hybrid force/position robot control. Proceedings of the IEEE
international conference on industrial technology, Bangkok, Thailand,
pp. 376-381.
Yao J, Chai Y (2007). Swing-up control of double inverted pendulum
based on adaptive neuro-fuzzy inference system. Fourth international
conference on fuzzy systems and knowledge discovery, Haikou,
Hainan, China, pp. 195-198.
Zhou C, Jagannathan K (1996). Adaptive network based fuzzy control
of a dynamic biped walking robot. Proceedings of the IEEE
international joint symposia on intelligence and systems, Maryland,
USA, 109-116.
Zhou YS, Lai LY (2000). Optimal design for fuzzy controllers by genetic
algorithms. IEEE Trans. Ind. Appl., 36(1): 93-97.