Hybrid Systems using Fuzzy, NN and GA (Soft Computing)

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
Hybrid Systems – Integration of NN, GA and FS : Course Lecture 41 – 42, notes, slides
www.myreaders.info/ , RC Chakraborty, e-mail rcchak@gmail.com , Dec. 01, 2010
http://www.myreaders.info/html/soft_computing.html
Hybrid Systems
Integration of Neural Network,
Fuzzy Logic & Genetic Algorithm
Soft Computing
www.myreaders.info
Return to Website
Hybrid systems, topic : Integration of neural networks, fuzzy logic
and genetic algorithms; Hybrid systems - sequential, auxiliar, and
embedded; Neuro-Fuzzy hybrid - integration of NN and FL; Neuro-
Genetic hybrids - integration of GAs and NNs ; Fuzzy-Genetic
hybrids - integration of FL and GAs. Genetic Algorithms Based
Back Propagation Networks : hybridization of BPN and GAs;
Genetic algorithms based techniques for determining weights in
a BPN - coding, weight extraction, fitness function algorithm,
reproduction of offspring, selection of parent chromosomes,
convergence. Fuzzy back propagation networks : LR-type fuzzy
numbers, operations on LR-type fuzzy numbers; Fuzzy neuron;
Architecture of fuzzy BP. Fuzzy associative memories : example
of FAM Model of washing machine - variables, operations,
representation, defuzzification. Simplified fuzzy ARTMAP :
supervised ARTMAP system, comparing ARTMAP with back-
propagation networks.

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
Hybrid Systems
Integration of Neural Network,
Fuzzy Logic & Genetic Algorithm
Soft Computing
Topics
(Lectures 41, 42 2 hours) Slides
1. Integration of Neural Networks, Fuzzy Logic and Genetic Algorithms
Hybrid systems : Sequential, Auxiliar, Embedded; Neuro-Fuzzy Hybrid :
Integration of NN and FL; Neuro-Genetic Hybrids : Integration of GAs
and NNs ; Fuzzy-Genetic Hybrids : Integration of FL and GAs; Typical
Hybrid systems.
03-13
2. Genetic Algorithms Based Back Propagation Networks
Hybridization of BPN and GAs; GA based techniques for determining
weights in a BPN : Coding, Weight extraction, Fitness function algorithm,
Reproduction of offspring, Selection of parent chromosomes,
Convergence.
14-25
3. Fuzzy Back Propagation Networks
LR-type Fuzzy numbers; Operations on LR-type Fuzzy Numbers; Fuzzy
Neuron; Architecture of Fuzzy BP.
26-32
4. Fuzzy Associative Memories
Example : FAM Model of Washing Machine - Variables, Operations,
Representation, Defuzzification.
33-37
5. Simplified Fuzzy ARTMAP
Supervised ARTMAP system, Comparing ARTMAP with Back-Propagation
Networks.
38-40
6. References 41
02

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
Hybrid Systems
Integration of NN FL GA
What is Hybridization ?
• Hybrid systems employ more than one technology to solve a problem.
• Hybridization of technologies can have pitfalls and therefore need to
be done with care.
• If one technology can solve a problem then a hybrid technology
ought to be used only if its application results in a better solution.
• Hybrid systems have been classified as :
− Sequential hybrid system: the technologies are used in pipelining
fashion;
− Auxiliary hybrid system: the one technology calls the other technology
as subroutine;
− Embedded hybrid system : the technologies participating appear to be
fused totally.
• Hybridization of fuzzy logic, neural networks, genetic algorithms has led
to creation of a perspective scientific trend known as soft computing.
− Neural networks mimic our ability to adapt to circumstances and learn
from past experience,
− Fuzzy logic addresses the imprecision or vagueness in input and output,
− Genetic algorithms are inspired by biological evolution, can systemize
random search and reach to optimum characteristics.
• Each of these technologies have provided efficient solution to wide range of
problems belonging to different domains. However, each of these
technologies has advantages and disadvantages. It is therefore appropriate
that Hybridization of these three technologies are done so as to over
come the weakness of one with the strength of other.
03

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
SC – Hybrid Systems - Introduction
1. Introduction :
Hybridization - Integration of NN , FL , and GA
Fuzzy logic, Neural networks and Genetic algorithms are soft computing
methods which are inspired by biological computational processes and
nature's problem solving strategies.
Neural Networks (NNs) are highly simplified model of human nervous system
which mimic our ability to adapt to circumstances and learn from past
experience. Neural Networks systems are represented by different architectures
like single and multilayer feed forward network. The networks offers back
proposition generalization, associative memory and adaptive resonance theory.
Fuzzy logic addresses the imprecision or vagueness in input and output
description of the system. The sets have no crisp boundaries and provide a
gradual transition among the members and non-members of the set elements.
Genetic algorithms are inspired by biological evolution, can systemize random
search and reach to optimum characteristics.
Each of these technologies have provided efficient solution to wide range
of problems belonging to different domains. However, each of these
technologies suffer from advantages and disadvantages.
It is therefore appropriate that Hybridization of these three technologies are
done so as to over come the weakness of one with the strength of other.
04

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
1.1 Hybrid Systems
Hybrid systems employ more than one technology to solve a problem.
Hybridization of technologies can have pitfalls and therefore need
to be done with care. If one technology can solve a problem then
a hybrid technology ought to be used only if its application results
in a better solution. Hybrid systems have been classified as
Sequential , Auxiliary and Embedded.
In Sequential hybrid system, the technologies are used in
pipelining fashion.
In Auxiliary hybrid system, one technology calls the other technology
as subroutine.
In Embedded hybrid system, the technologies participating appear to
be fused totally.
05

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Sequential Hybrid System
In Sequential hybrid system, the technologies are used in pipelining
fashion. Thus, one technology's output becomes another technology's
input and it goes on. However, this is one of the weakest form of
hybridization since an integrated combination of technologies is not
present.
Example: A Genetic algorithm preprocessor obtains the optimal
parameters for different instances of a problem and hands over the
preprocessed data to a neural network for further processing.
06

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Auxiliary Hybrid System
In Auxiliary hybrid system, one technology calls the other technology
as subroutine to process or manipulate information needed. The second
technology processes the information provided by the first and hands
it over for further use. This type of hybridization is better than the
sequential hybrids.
Example : A neuron-genetic system in which a neural network
employs a genetic algorithm to optimize its structural parameters
that defines its architecture.
07

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Embedded Hybrid System
In Embedded hybrid system, the technologies participating are
integrated in such a manner that they appear intertwined. The fusion
is so complete that it would appear that no technology can be used
without the others for solving the problem.
Example : A NN-FL hybrid system may have an NN which receives
fuzzy inputs, processes it and extracts fuzzy outputs as well.
08

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
1.2 Neural Networks, Fuzzy Logic, and Genetic Algorithms Hybrids
Neural Networks, Fuzzy Logic, and Genetic Algorithms are three
distinct technologies.
Each of these technologies has advantages and disadvantages. It is
therefore appropriate that hybridization of these three technologies are
done so as to over come the weakness of one with the strength
of other.
09

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Neuro-Fuzzy Hybrid
Neural Networks and Fuzzy logic represents two distinct methodologies to
deal with uncertainty. Each of these has its own merits and demerits.
Neural Networks :
− Merits : Neural Networks, can model complex nonlinear relationships
and are appropriately suited for classification phenomenon into
predetermined classes.
− Demerits : Neural Network's output, precision is often limited to least
squares errors; the training time required is quite large; the training
data has to be chosen over entire range where the variables are
expected to change.
Fuzzy logic :
− Merits : Fuzzy logic system, addresses the imprecision of inputs and
outputs defined by fuzzy sets and allow greater flexibility in
formulating detail system description.
Integration of NN and FL, called Neuro-Fuzzy systems, have the potential
to extend the capabilities of the systems beyond either of these two
technologies applied individually. The integrated systems have turned
out to be useful in :
− accomplishing mathematical relationships among many variables in a
complex dynamic process,
− performing mapping with some degree of imprecision, and
− controlling nonlinear systems to an extent not possible with conventional
linear control systems.
There are two ways to do hybridization :
− One, is to provide NNs with fuzzy capabilities, there by increasing the
network's expressiveness and flexibility to adapt to uncertain
environments.
− Second, is to apply neuronal learning capabilities to fuzzy systems so
that the fuzzy systems become more adaptive to changing
environments. This method is called NN driven fuzzy reasoning.
10

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Neuro-Genetic Hybrids
The Neural Networks and Genetic Algorithms represents two distinct
methodologies.
Neural Networks : can learn various tasks from examples, classify
phenomena and model nonlinear relationships.
Genetic Algorithms : have offered themselves as potential candidates
for the optimization of parameters of NN.
Integration of GAs and NNs has turned out to be useful.
− Genetically evolved nets have reported comparable results against their
conventional counterparts.
− The gradient descent learning algorithms have reported difficulties in
leaning the topology of the networks whose weights they optimize.
− GA based algorithms have provided encouraging results especially
with regard to face recognition, animal control, and others.
− Genetic algorithms encode the parameters of NNs as a string of
properties of the network, i.e. chromosomes. A large population of
chromosomes representing many possible parameters sets, for the
given NN, is generated.
− GA-NN is also known as GANN have the ability to locate the
neighborhood of the optimal solution quicker than other conventional
search strategies.
− The drawbacks of GANN algorithms are : large amount of memory
required to handle and manipulate chromosomes for a given network;
the question is whether this problem scales as the size of the networks
become large.
11

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Fuzzy-Genetic Hybrids
Fuzzy systems have been integrated with GAs.
The fuzzy systems like NNs (feed forward) are universal approximator
in the sense that they exhibit the capability to approximate general
nonlinear functions to any desired degree of accuracy.
The adjustments of system parameters called for in the process, so
that the system output matches the training data, have been tackled using
GAs. Several parameters which a fuzzy system is involved with like
input/output variables and the membership function that define the
fuzzy systems, have been optimized using GAs.
12

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
1.3 Typical Hybrid Systems
The Systems considered are listed below.
1. Genetic algorithm based back propagation network
(Neuro Genetic Hybrid)
2. Fuzzy back propagation network
(Neuro – Fuzzy Hybrid with Multilayer Feed forward Network as the
host architecture)
(Neuro – Fuzzy Hybrid with Recurrent Network as the host architecture)
4. Fuzzy Associative Memory
( Neuro – Fuzzy Hybrid with single layer Feed forward architecture)
5. Fuzzy logic controlled Genetic algorithm
(Fuzzy – Genetic Hybrid)
13

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
SC – Hybrid Systems – GA based BPN
2. Genetic Algorithm (GA) based Back Propagation Network (BPN)
Neural networks (NNs) are the adaptive system that changes its structure
based on external or internal information that flows through the network.
Neural network solve problems by self-learning and self-organizing.
Back Propagation Network (BPN) is a method of training multi-layer neural
networks. Here learning occurs during this training phase.
The steps involved are:
− The pattern of activation arriving at the output layer is compared with the
correct output pattern to calculate an error signal.
− The error signal is then back-propagated from output to input for
adjusting the weights in each layer of the BPN.
− The Back-Propagation searches on the error surface using gradient descent
method to minimize error E = 1/2 Σ ( T j – O j )2
where T j is target output
and O j is the calculated output by the network.
Limitations of BPN :
− BPN can recognize patterns similar to those they have learnt, but do not
have the ability to recognize new patterns.
− BPN must be sufficiently trained to extract enough general features
applicable to both seen and unseen; over training to network may have
undesired effects.
14

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
[ continued from previous slide ]
Genetic Algorithms (GAs) are adaptive search and optimization algorithms,
mimic the principles of nature.
− GAs are different form traditional search and
− Optimization exhibit simplicity, ease of operation, minimal requirements,
and global perspective.
Hybridization of BPN and GAs
− The BPN determines its weight based on gradient search technique and
therefore it may encounter a local minima problem.
− GAs do not guarantee to find global optimum solution, but are good in
finding quickly good acceptable solution.
− Therefore, hybridization of BPN and GAs are expected to provide many
advantages compare to what they alone can.
The GA based techniques for determining weights in a BPN are explained next.
15

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
2.1 GA based techniques for determining weights in a BPN
Genetic algorithms work with population of individual strings.
The steps involved in GAs are:
− each individual string represent a possible solution of the problem
considered,
− each individual string is assigned a fitness value,
− high fit individuals participate in reproduction, yields new strings as
offspring and they share some features with each parents,
− low fit individuals are kept out from reproduction and so die,
− a whole new population of possible solutions to the problem is
generated by selecting high fit individuals from current generation,
− this new generation contains characteristics which are better than
their ancestors,
− processing this way after many generation, the entire population
inherits the best and fit solution.
However, before a GA is executed :
− a suitable coding for the problem is devised,
− a fitness function is formulated,
− parents have to be selected for reproduction and crossover to
generate offspring.
All these aspects of GAs for determining weights of BPN are illustrated
in next few slides.
16

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Coding
Assume a BPN configuration ℓ - m – n where
− ℓ is input , m is hidden and n is output neurons.
− number of weights to be determined are (ℓ + n) m.
− each weight (gene) is a real number.
− assume number of digits (gene length) in weight are d .
− a string S represents weight matrices of input-hidden and the hidden-
output layers in a linear form arranged as row-major or column-major
selected.
− population size is the randomly generated initial population of p
chromosomes.
Example :
Consider a BPN configuration ℓ - m – n where ℓ = 2 is input , m = 2 is
hidden and n = 2 is output neuron.
Input neuron Hidden neurons output neurons
Input layer Hidden layer output layer
Fig. BPN with 2 – 2 - 2
− number of weights is (ℓ + n) m
= ( 2 + 2) . 2 = 8
− each weight is real number and
assume number of digits in
weight are d = 5
− string S representing
chromosome of weights is 8 x 5
= 40 in length
− Choose a population size p = 40
ie choose 40 chromosomes
Gene
← k=0 →
Gene
← k=1 →
Gene
← k=2 →
Gene
← k=3 →
Gene
← k=4 →
Gene
← k=5 →
Gene
← k=6 →
Gene
← k=7 →
84321 46234 78901 32104 42689 63421 46421 87640
Chromosome
32478 76510 02461 84753 64321 14261 87654 12367
Chromosome
Fig. Some randomly generated chromosome made of 8 genes
representing 8 weights for BPN
17
1
2
1 1
2 2
W11
Inputs
W12
W21
W22
V11
V22
V12
V21
Outputs

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Weight Extraction
Extract weights from each chromosomes, later to determine the fitness
values.
Let x1 , x2 , . . . . x d , . . . . x L represent a chromosome and
Let xkd+1 , xkd+2 , . . x(k + 1)d represent k
th
gene (k ≥ 0) in the chromosomes.
The actual weight wk is given by
xkd+2 10d-2
+ xkd +3 10d-3
+ . . . + x(k + 1)d , if 5 ≤ xkd +1 ≤ 9
10d-2
wk =
xkd +2 10d-2
+ xkd +3 10d-3
+ . . . + x(k + 1)d , if 0 ≤ xkd +1 < 5
10d-2
Example : [Ref Fig. BPN previous slide]
The Chromosomes are stated in the Fig. The weights extracted from all
the eight genes are :
■ Gene 0 : 84321 ,
Here we have, k = 0 , d = 5 , and xkd +1 is x1 such that
5 ≤ x1 = 8 ≤ 9. Hence, the weight extracted is
4 x 103
+ 3 x 102
+ 2 x 10 + 1
103
■ Gene 1 : 46234 ,
Here we have, k = 1 , d = 5 , and xkd +1 is x6 such that
0 ≤ x6 = 4 ≤ 5. Hence, the weight extracted is
6 x 103
+ 2 x 102
+ 3 x 10 + 4
103
■ Similarly for the remaining genes
Gene 2 : 78901 yields W2 = + 8.901
Gene 3 : 32104 yields W3 = − 2.104
Gene 4 : 42689 yields W4 = − 2.689
Gene 5 : 63421 yields W5 = + 3.421
Gene 6 : 46421 yields W6 = − 6.421
Gene 7 : 87640 yields W7 = + 7.640
18
+
W0 = + = +4.321
W1 = − = − 6.234

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Fitness Function :
A fitness is devised for each problem.
Example :
The matrix on the right, represents a set of input I
and output T for problem P to be solved.
Generate initial population P0 of size p = 40.
(I11 , I21) (T11 , T21)
(I12 , I22) (T12 , T22)
(I13 , I23) (T13 , T23)
Let C0
1 , C0
1 , . . . , C0
40 represent the 40 chromosomes.
Let , , . . . . be the weight sets extracted, using the Eq.
in the previous slides, from each of the chromosome C0
i , i = 1, 2, . . . , 40 .
Let , , be the calculated outputs of BPN.
Compute root mean square error :
E1 = (T11 – O11)2
+ (T21 – O21)2
,
E2 = (T12 – O12)2
+ (T22 – O22)2
E3 = (T13 – O13)2
+ (T23 – O23)2
The root mean square of error is
E = [(E1 + E2 + E3) / 3 ] 1/2
Compute Fitness F1 :
The fitness F1 for the chromosome C0
1 is given by
F1 = 1 / E .
Similarly, find the fitness F2 for the chromosome C
0
2 and
so on the fitness Fn for the chromosome C0
n
19
w0
2
w0
1 w0
40
o 0
2
o0
1 o 0
3

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
[ continued from previous slide fitness fuction]
Algorithm
{
Let ( , ) , i = 1 , 2 , . . . , N represents the input-output pairs of the
problem to be solved by BPN with configuration ℓ - m – n ; where
= (I1i , I2i , , . . . , I ℓ i ) and
= (T1i , T2i , , . . . , Tn i )
For each chromosome C i , i = 1 , 2 , . . . , p belonging to current the
population P i whose size is p
{
Extract weights form C i using Eq. 2.1 in previous slide;
Keeping as a fixed weight, train the BPN for the N input instances;
Calculate error E i for each of the input instances using the formula below
E i = ( T j i – O j i )
2
where is the output vector calculated by BPN;
Find the root mean square E of the errors E i , i = 1 , 2 , . . . , N
i.e. E = ( ( E i ) / N ) 1/2
Calculate the Fitness value F i for each of the individual string of the
population as F i = 1 / E
}
Output F i for each C i , i = 1 , 2 , . . . , p ;
}
20
Ii Ti
I
Ti
w i
w i
Σ
j
O i
Σ
i

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
[ continued from previous slide - Fitness Function ]
Thus the Fitness values Fi for all chromosomes in the initial
population are computed. The population size is p = 40, so F i , i = 1
, 2 , . . , 40 are computed.
A schematic for the computation of fitness values is illustrated below.
Fig. Computation of Fitness values for the population
21
C0
1
C0
2
----
----
C0
40
F i =1/E
---
---
w0
1
w0
2
W
0
40 Fitness
Values
Initial
Population of
Chromosomes
Extracted
weight sets
Training BPN
Extract Input Output
Error E
weights
weights
Compute
Fitness

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Reproduction of Offspring
Before the parent chromosomes reproduce offspring :
First, form a mating pool by excluding that chromosome C ℓ with least
fitness F
min
and then replacing it with a duplicate copy of C k with
highest fitness F max
;
i.e., the best fit individuals have multiple copies while worst fit
individuals die off.
Having formed the mating pool, select parent pair at random.
Chromosomes of respective pairs are combined using crossover
operator. Fig. below shows :
− two parent chromosomes Pa and Pb,
− the two point crossover,
− exchange of gene segments by the parent pairs, and
− the offspring Oa and Ob are produced.
Pa Pb
Oa Ob
Fig. Two – point crossover operator
22
Parent
Chromosomes
Offspring
A B
B A
Crossover
Point 1
Crossover
Point 1
Crossover
Point 1
Crossover
Point 1

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
[ continued from previous slide - Reproduction ]
Example :
− Consider the initial population of chromosomes P0 generated, with
their fitness value F i , where i = 1 , 2 , . . , 40 ,
− Let F
max
= Fk be maximum and F
min
= F ℓ be minimum fitness value
for 1 ≤ ℓ , k ≤ 40 where ℓ ≠ k
− Replace all chromosomes having fitness value F min
with copies of
chromosomes having fitness value F
max
Fig. below illustrates the Initial population of chromosomes and the
formation of the mating pool.
Initial population P0 Mating pool
Fig. Formation of Mating pool
F min
is replaced by F max
23
C0
1 F1
C0
2 F2
C0
k Fk
C0
ℓ F ℓ
----
----
C0
40 F40
C0
1 F1
C0
2 F2
C0
k Fk
C0
ℓ F max
----
----
C0
40 F40
Max Fitness
value F
max
Min Fitness
value F
min
Chromosomes
C0
1 to C0
40

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Selection of Parent Chromosomes
The previous slide illustrated Reproduction of the Offspring.
Here, sample "Selection Of Parents" for the "Two Points Crossover" operator
to produce Offspring Chromosomes are illustrated.
Fig. Random Selection of Parent Chromosomes
The Crossover Points of the Chromosomes are randomly chosen for each
parent pairs as shown in the Fig. below.
Fig. Randomly chosen Crossover points of Parent Chromosomes
The Genes are exchanged for Mutation as shown in the Fig. below.
Fig. New population P1 after application of two point Crossover operator
Thus new population P1 is created comprising 40 Chromosomes which
are the Offspring of the earlier population generation P0 .
24
Chromosomes - Mating Pool
Selected Parent Pairs
C1
1 C1
k C1
ℓ C1
40
C1
2
Chromosomes -Mating Pool
Crossover
points
Selected Parent Pairs
C1
1 C1
k C1
ℓ C1
40
C1
2
Chromosomes -Mating Pool
C1
1 C1
k C1
ℓ C1
40
C1
2
New Population P1
C1
1 C1
k C1
ℓ C1
40
C1
2

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Convergence
For any problem, if GA is correctly implemented, the population evolves
over successive generations with fitness value increasing towards the
global optimum.
Convergence is the progression towards increasing uniformity.
A population is said to have converged when 95% of the individuals
constituting the population share the same fitness value.
Example :
Let a population P1 undergoes the process of selection, reproduction,
and crossover.
− the fitness values for the chromosomes in P1 are computed.
− the best individuals replicated and the reproduction carried out using
two-point crossover operators form the next generation P2 of the
chromosomes.
− the process of generation proceeds until at one stage 95% of the
chromosomes in the population Pi converge to the same fitness value.
− at that stage, the weights extracted from the population Pi are the
final weights to be used by BPN.
25

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
SC – Hybrid Systems – Fuzzy BPN
3. Fuzzy Back Propagation Network
Neural Networks and Fuzzy logic (NN-FL) represents two distinct methodologies
and the integration of NN and FL is called Neuro-Fuzzy systems.
Back Propagation Network (BPN) is a method of training multi-layer neural
networks where learning occurs during this training phase.
Fuzzy Back Propagation Network (Fuzzy-BPN) is a hybrid architecture. It is,
Hybridization of BPN by incorporating fuzzy logic.
Fuzzy-BPN architecture, maps fuzzy inputs to crisp outputs. Here, the
Neurons uses LR-type fuzzy numbers.
The Fuzzy-Neuron structure, the architecture of fuzzy BP, its learning
mechanism and algorithms are illustrated in next few slides.
26

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
3.1 LR-type Fuzzy Numbers
The LR-type fuzzy number are special type of representation of fuzzy
numbers. They introduce functions called L and R.
• Definition
A fuzzy member is of L-R type if and only if
where L is a left reference
R is a right reference,
m , is called mean of is a real number,
α , β are left and right spreads respectively.
µ is the membership function of fuzzy member
The functions L and R are defined as follows:
LR-type fuzzy number can be represented as (m, α, β) LR shown below.
1
Member ship
deg
00 α m, β x
Fig. A triangular fuzzy number (m, α, β).
Note : If α and β are both zero, then L-R type function indicates a
crisp value. The choice of L and R functions is specific to problem.
27
L for x ≤ m , α > 0
=
R for x ≤ m , β > 0
µ (x)
M
~
m – x
α
m – x
β
L = max ( 0 , 1 - )
R = max ( 0 , 1 - )
m – x
α
m – x
α
m – x
α
m – x
α
µ (x)
M
~
M
~
M
~
M
~
M
~
M
~

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Operations on LR-type Fuzzy Numbers
Let = (m, α , β) LR and = (n, γ , δ) LR be two L R-type fuzzy
numbers. The basic operations are
■ Addition
(m, α , β) LR (n, γ , δ) LR = (m + n, α + γ , β + δ ) LR
■ Substraction
(m, α , β) LR (n, γ , δ) LR = (m - n, α + δ , β + γ ) LR
■ Multiplicaion
(m, α , β) LR (n, γ , δ) LR = (mn , mγ + nα , mδ + nβ) LR for m≥0 , n≥0
(m, α , β) LR (n, γ , δ) LR = (mn , mα - mδ , nβ - mγ) RL for m<0 , n≥0
(m, α , β) LR (n, γ , δ) LR = (mn , - nβ - mδ , -nα - mγ) LR for m<0 ,
n<0
■ Scalar Multiplicaion
λ*(m, α , β) LR = (λm, λα , λβ) LR , ∀λ ≥ 0 , λ ∈ R
λ*(m, α , β) LR = (λm, -λα , -λβ) RL , ∀λ < 0 , λ ∈ R
28
M
~
N
~

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Fuzzy Neuron
The fuzzy neuron is the basic element of Fuzzy BP network. Fig. below
shows the architecture of the fuzzy neuron.
Fig Fuzzy Neuron j
Given input vector
and weight vector
The fuzzy neuron computes the crisp output given by
O = f (NET) = f ( CE ( . )) where = (1, 0, 0) is the bias.
Here, the fuzzy weighted summation is given by
is first computed and
is computed next
The function CE is the centroid of triangular fuzzy number, that has
m as mean and α , β as left and right spreads explained before, can
be treated as defuzzification operation, which maps fuzzy weighted
summation to crisp value.
If is the fuzzy weighted summation
Then function CE is given by
The function f is a sigmoidal function that performs nonlinear mapping
between the input and output. The function f is obtained as :
f (NET) = 1 / ( 1 + exp ( - NET ) ) = O is final crisp output value.
29
O j
Function
CE
∑ f
I1
~
I2
~
I3
~
In-1
~
In
~
Net j
~
Wn
~
W2
~
W3
~
Wn-1
~
W1
~
Σ
i=1
n
, , , . .
I
~
In
~
I2
~
I1
~
=
, , , . .
W
~
wn
~
w2
~
w1
~
=
Wi
~
Ii
~
I0
~
Wi
~
Ii
~
Σ
i=0
n
net
~
= ▪
net
~
NET CE
= ( )
netm
~
net
~
= ( )
netα
~
netβ
~
, ,
=
( )
CE net
~
netm
~
( )
netα
~
netβ
~
, ,
CE netm
~
netβ
~
netα
~
= –
( )
+1/3 = NET

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
[ continued from previous slide – Fuzzy Neuron ]
Note :
In the fuzzy neuron, both input vector and weight vector are
represented by triangular LR-type fuzzy numbers.
For input vector the input component is
represented by the LR-type fuzzy number , , .
Similarly, for the weight vector the weight vector
component is represented as , , .
30
In
~
wn
~
, , , . .
I
~
In
~
I2
~
I1
~
= Ii
~
I m i
~
I α i
~
I β i
~
, , , . .
W
~
wn
~
w2
~
w1
~
=
wi
~
w m i
~
w α i
~
w β i
~

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Architecture of Fuzzy BP
Fuzzy Back Propagation Network (BP) is a 3-layered feed forward
architecture. The 3 layers are: input layer, hidden layer and output layer.
Considering a configuration of ℓ-input neurons, m-hidden neurons and
n-output neurons, the architecture of Fuzzy BP is shown below.
.
Fig. Three layer Fuzzy BP architecture.
Let , for p = 1, 2, . . , N, be the pth
pattern
among N input patterns that Fuzzy BP needs to be trained.
Here, indicates the i th
component of input pattern p and is an LR-
type triangular fuzzy number, i.e.,
− Let be the output value of i
th
input neuron.
− Let O'pj and O'pk are jth
and kth
crisp defuzzification outputs of
the hidden and output layer neurons respectively.
− Let Wij is the fuzzy connection weight between i th
input node and
j
th
hidden node.
− Let Vjk is the fuzzy connection weight between j
th
hidden node and
k
th
output node.
[Continued in next slide]
31
, , , . .
Ip
~
Ipℓ
~
Ip2
~
Ip1
~
=
1
1 1
i j k
ℓ n
m
I"
p1
Op1
Ip1 V11
I'p1
O'p1 O"
p1
W11
I"
pk
Opj
Ipi I'pj O'pj O"
pk
I"
pn
Opℓ O'pm
I'pm
Ipℓ O"
pn
Wij Vjk
V1k
W1j
Wℓm Vmn
~
~
~
~
~
~
Ipi
~
~
, , , . .
Ip
~
Ipβℓ
Ip2
~
Ip1
~
=
Opi
~

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
[ continued from previous slide – Architecture of Fuzzy BP]
The computations carried out by each layer are as follows:
Input neurons:
= , i = 1 , 2 , . . . , ℓ .
Hidden neurons:
O' pj = f ( NET pj ) , i = 1 , 2 , . . . , m .
where NET pj = C E ( Wij O' pi )
Out neurons:
O" pk = f ( NET pk ) , i = 1 , 2 , . . . , n . ,
where NET pk = C E ( Vjk O' pj )
32
Σ
i=0
ℓ
Σ
j=0
m
Opi
~
Ipi
~

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
SC – Hybrid Systems – Fuzzy AM
4. Fuzzy Associative Memory
A fuzzy logic system contains the sets used to categorize input data (i.e.,
fuzzification), the decision rules that are applied to each set, and then a
way of generating an output from the rule results (i.e., defuzzification).
In the fuzzification stage, a data point is assigned a degree of
membership (DOM) determined by a membership function. The member-
ship function is often a triangular function centered at a given point.
The Defuzzification is the name for a procedure to produce a real
(non-fuzzy) output .
Associative Memory is a type of memory with a generalized addressing
method. The address is not the same as the data location, as in
traditional memory. An associative memory system stores mappings
of specific input representations to specific output representations.
Associative memory allows a fuzzy rule base to be stored. The inputs are
the degrees of membership, and the outputs are the fuzzy system’s output.
Fuzzy Associative Memory (FAM) consists of a single-layer feed-forward
fuzzy neural network that stores fuzzy rules "If x is Xk then y is Yk" by
means of a fuzzy associative matrix.
FAM has many applications; one such application is modeling the operations
of washing machine.
33

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
• Example : Washing Machine (FAM Model)
For a washing machine , the input/output variables are :
Output variable : washing time (T) , depends upon two input variables.
Input variables are : weight of clothes (X) and stream of water (Y).
These variables have three different degree of variations as :
small (S), medium (M), and large (L) .
These three variables X , Y, and T, are defined below showing their
membership functions µX , µY and µT .
■ Clothes weight is X,
− range is from 0 to 10 and
− the unit is kilogram (k.g).
Weight (X)
■ Stream of water is Y
− the unit is liter per minute (liters/min)
Stream (Y)
■ Washing time is T
− the unit is minutes (min.)
Washing time (T)
34
S M L
2.5 5.0 7.5 10
0.0
0.0
0.2
0.4
0.6
0.8
1.0
µX
16 40 64 80
0.0
0.0
0.2
0.4
0.6
0.8
1.0
µY
25 50 75 100
0.0
0.0
0.2
0.4
0.6
0.8
1.0
µT
S M L
S M L

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
[ continued from previous slide – Model of Washing Machine]
The problem indicates, that there are two inputs and one-output
variables. The inference engineer is constructed based on fuzzy rule :
“ If < input variable > AND < input variable >
THEN < output variable >”
According to the above fuzzy rule, the Fuzzy Associative Memory
(FSM) of X, Y, and T variables are listed in the Table below.
Weight (X)
Washing time (T)
S M L
S M L L
M S M L
Stream (Y)
L S S L
Table 1. Fuzzy associative memory (FSM) of Washing Machine
■ Operations : To wash the clothes
− Turn on the power,
− The machine automatically detects the weight of the clothes
as (X) = 3.2 K.g. ,
− The machine adjusts the water stream (Y) to 32 liter/min.,
35

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
■ Fuzzy Representation :
The fuzzy sets representation, while X = 3.2 Kg and Y = 32
liter/min., according to the membership functions, are as follows:
The fuzzy set of X3.2 Kg = { 0.8/S, 0.2/M, 0/L }
The fuzzy set of Y32 liters/min. = { 0.4/S, 0.8/M, 0/L }
Washing time (T)
Fig. Simulated Fuzzy set representation of washing machine
36
25 50 75 100
0.0
0.0
0.2
0.4
0.6
0.8
1.0
µT
20 35 60

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
■ Defuzzification
The real washing time is defuzzied by the Center of gravity (COG)
defuzzification formula. The washing time is calculated as :
Z COG = µc (Z j ) Z j / µc (Z j ) where
j = 1, . . . , n , is the number of quantization levels of the output,
Z j is the control output at the quantization level j ,
µc (Z j ) represents its membership value in the output fuzzy set.
Referring to Fig in the previous slide and the formula for COG, we get
the fuzzy set of the washing time as w = { 0.8/20, 0.4/35, 0.2/60 }
The calculated washing time using COG formula T = 41.025 min.
37
Σ
j=1
n
Σ
j=1
n

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
SC – Hybrid Systems – Fuzzy ART
ART is a neural network topology whose dynamics are based on Adaptive
Resonance Theory (ART). ART networks follow both supervised and
unsupervised algorithms.
− The Unsupervised ARTs are similar to many iterative clustering
algorithms where "nearest" and "closer" are modified slightly by
introducing the concept of "resonance". Resonance is just a matter of
being within a certain threshold of a second similarity measure.
− The Supervised ART algorithms that are named with the suffix "MAP", as
ARTMAP. Here the algorithms cluster both the inputs and targets and
associate two sets of clusters.
The basic ART system is an unsupervised learning model.
The ART systems have many variations : ART1, ART2, Fuzzy ART, ARTMAP.
ART1: The simplest variety of ART networks, accepting only binary inputs.
ART2 : It extends network capabilities to support continuous inputs.
ARTMAP : Also known as Predictive ART. It combines two slightly
modified ART-1 or ART-2 units into a supervised learning structure. Here,
the first unit takes the input data and the second unit takes the correct
output data, then used to make the minimum possible adjustment of the
vigilance parameter in the first unit in order to make the correct
classification.
The Fuzzy ARTMAP model is fuzzy logic based computations incorporated
in the ARTMAP model.
Fuzzy ARTMAP is neural network architecture for conducting supervised
learning in a multidimensional setting. When Fuzzy ARTMAP is used on a
learning problem, it is trained till it correctly classifies all training data. This
feature causes Fuzzy ARTMAP to ‘over-fit’ some data sets, especially those
in which the underlying pattern has to overlap. To avoid the problem of
‘over-fitting’ we must allow for error in the training process.
38

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
SC – Hybrid Systems – Fuzzy ART
• Supervised ARTMAP System
ARTMAP is also known as predictive ART. The Fig. below shows a
supervised ARTMAP system. Here, two ART modules are linked by an
inter-ART module called the Map Field. The Map Field forms predictive
associations between categories of the ART modules and realizes a match
tracking rule. If ARTa and ARTb are disconnected then each module
would be of self-organize category, groupings their respective input sets.
Fig. Supervised ARTMAP system
In supervised mode, the mappings are learned between input vectors
a and b. A familiar example of supervised neural networks are
feed-forward networks with back-propagation of errors.
39
ART a
ART b
MAP Field
Map Field
Orienting
Subsystem
Map Field
Gain
Control
Match
Tracking
Training
b
a

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
Sc – Hybrid Systems – Fuzzy ART
• Comparing ARTMAP with Back-Propagation Networks
ARTMAP networks are self-stabilizing, while in BP networks the new
information gradually washes away old information. A consequence of
this is that a BP network has separate training and performance
phases while ARTMAP systems perform and learn at the same time
− ARTMAP networks are designed to work in real-time, while BP networks
are typically designed to work off-line, at least during their training
phase.
− ARTMAP systems can learn both in a fast as well as in a slow match
configuration, while, the BP networks can only learn in slow mismatch
configuration. This means that an ARTMAP system learns, or adapts its
weights, only when the input matches an established category, while
BP networks learn when the input does not match an established
category.
− In BP networks there is always a danger of the system getting
trapped in a local minimum while this is impossible for ART systems.
However, the systems based on ART modules learning may depend
upon the ordering of the input patterns.
40

R
C
C
h
a
k
r
a
b
o
r
t
y
,
w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
SC – Hybrid Systems
6. References : Textbooks
1. "Neural Network, Fuzzy Logic, and Genetic Algorithms - Synthesis and
Applications", by S. Rajasekaran and G.A. Vijayalaksmi Pai, (2005), Prentice Hall,
Chapter 10-15, page 297-435.
2. “Soft Computing and Intelligent Systems - Theory and Application”, by Naresh K.
Sinha and Madan M. Gupta (2000), Academic Press, Chapter 1-25, page 1-625.
3. "Soft Computing and Intelligent Systems Design - Theory, Tools and Applications",
by Fakhreddine karray and Clarence de Silva (2004), Addison Wesley, chapter 7,
page 337-361.
4. “Neuro-Fuzzy and Soft Computing: A Computational Approach to Learning and
Machine Intelligence” by J. S. R. Jang, C. T. Sun, and E. Mizutani, (1996),
Prentice Hall, Chapter 17-21, page 453-567.
5. "Fuzzy Logic: Intelligence, Control, and Information", by John Yen, Reza Langari,
(1999 ), Prentice Hall, Chapter 15-17, page 425-500.
6. "Fuzzy Logic and Neuro Fuzzy Applications Explained", by Constantin Von Altrock,
(1995), Prentice Hall, Chapter 4, page 63-79.
7. Related documents from open source, mainly internet. An exhaustive list is
being prepared for inclusion at a later date.
41

Hybrid Systems using Fuzzy, NN and GA (Soft Computing)

Recommended

More Related Content

What's hot (20)

Similar to Hybrid Systems using Fuzzy, NN and GA (Soft Computing) (20)

More from Amit Kumar Rathi (20)

Recently uploaded (20)

Hybrid Systems using Fuzzy, NN and GA (Soft Computing)