The solution of problem of parameterization of the proximity function in ace using genetic algorithm

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
_______________________________________________________________________________________
Volume: 04 Issue: 12 | Dec-2015, Available @ http://www.ijret.org 100
THE SOLUTION OF PROBLEM OF PARAMETERIZATION OF THE
PROXIMITY FUNCTION IN ACE USING GENETIC ALGORITHM
Khamroev Alisher
Senior researcher, Centre for development software and hardware program complexes at the Tashkent University of
Information Technologies, Tashkent, Uzbekistan
Abstract
In this work, a new approach for defining the value of the proximity function, which is carried out in the second step of the
Algorithms for Calculating Estimates (ACE) in the area of Pattern Recognition, is presented. The value of the proximity function
is defined as a part of corresponding features of two objects. The main attention is paid to essential features of the polytypic in a
given training set. One of the important problems of the ACE is to compare the values of fuzzy attributes. The main idea of this
approach is considering the proximity the corresponding quantitative and qualitative features together. Here a complexity of
comparing the qualitative features and an approach of overcoming such complexity are considered. Such features include the
features with fuzzy values. The membership function of fuzzy set theory is used for determining membership degrees of the feature
values describing with linguistic values for improve the quality of ACE. The steps of the algorithm for transfer the results is
obtained from the comparison of the two values of fuzzy feature by using membership function to the proximity function. The
membership function with two parameters (b and c) is used. For defining optimal values of these parameters evolutionary
algorithms for solving optimization problems are used, one of them is Genetic algorithm. By using genetic algorithm initial
parameters’ values of the membership function are generated and transmitted to the proximity function. The ACE is run and value
of functional quality is defined during the training process with given training set. If the value of the functional quality is not
sufficiently high than the values obtained by Genetic algorithm, these values are regenerated using special operators (selection,
crossover, mutation) of the Genetic algorithm. The algorithm for selection optimal values of the parameters of the membership
function using the Genetic algorithm is given.
Key Words: ACE, proximity function, Genetic algorithm, membership function, parameters, operators.
---------------------------------------------------------------------***---------------------------------------------------------------------
1. INTRODUCTION
Problem of recognition is connected with the study of
human creative activity, the study of how the processes of
making a decision on solving various problems, how targets
are selected and concepts are formulated occur in the human
brain. These questions lie in sphere the so-called heuristic
principles and methods for solving problems. The task of
recognition of objects of different physical nature is that in
terms of set of features, which show an unknown object
from a defined class, it is needed to determine to what
specific object out of this class this particular set of features
corresponds to.
For the solution of the problem of recognition based on the
precedents many methods and algorithms were developed.
One of such recognition algorithms are algorithms based on
the principle of partial precedent (the algorithms for
calculation estimates - ACE), developed and improved
under the supervision of Yu.I. Zhuravlev [1, 2]. At the core
of these models is the idea of finding the right partial
precedents in features descriptions source data (informative
fragments of feature values or representative sets).
2. STATEMENT OF THE PROBLEM
Let there be a general set of }{S which consists of original
data (objects) MS j  , mj ,...,2,1 . Usually, naturally the
objects jS are given with polytypic features niXi ,...2,1,  ,
)...,,,( 21 njjjj xxxS  , iij Xx  . The elements (values of
features) iX make up an alphabetic features, for example,
they can be expressed in terms of “yes/no”,
“yes/no/unknown”, the numerical values, linguistic values,
the values of a set of possible options, etc. The task of
recognition – is the classification of source data to a specific
class using a selection of essential features, which
characterize the data, out of the total weight of the non-
essential data.
2.1 The Algorithms for Calculating Estimates
problem
It is known [1], that in the algorithms, the calculation of
recognition is based on comparing the recognizable object
with the standard objects by different sets of features using
the voting procedures. In most cases, the voting procedure is
conducted for quantitative features of objects. However, the
features of objects can be expressed by fuzzy values
(linguistic terms). In such cases, the proximity between
relevant two fuzzy features of objects should be considered.
For this we have to modify the second step of setting of

_______________________________________________________________________________________
ACE for the purpose of correctly calculating the value of the
proximity function ),( qj SSr . Here qS - control objects.
At the first stage, ACE is formed from full set of
( nXX ,...,1 ) under the possible collection of k featuring
k ,...,1 so-called system of support collection. Here k –
the length of the voting under the collection of n, which
represent some ~-part of the features of vector objects.
In the second stage of ACE the proximity function
)~,~( qj SSr  the value which determines measure of
proximity of the parts of rows (so-called ~ -parts) is
calculated. The proximity function )~,~( qj SSr  is
determined using following functions ),( qiji vv
xx , which
determine degree of proximity of features. These functions,
as a rule, are selected in accordance with the character of
used features.
The value of the proximity function can be set in the
following types (comparison: a strict-b-soft):
1)


 

otherwise,;0
~~;1
)~,~(
q
q
SS
SSr


2)









otherwise,;0
),(~;1
)~,~( 1


k
v
qiji
q
vv
xx
SSr
here ),(~
iqij xx - comparison function between the features
jiv
x and qiv
x ,  - threshold for ~ -parts.
We will consider some functions, which determine the
degree of proximity of the values of the corresponding
feature of two objects:
a) if the features are given in binary values, the following
function is entered:







iqij
iqij
iqij
xx
xx
xx
;0
,;1
),(
b) if the features are given in quantitative values, the
following function is entered which uses the parameter i -
thresholds:


 

otherwise,,0
,||если,1
),(
iiqij
iqij
xx
xx


c) if the features are given in fuzzy values, fuzzy terms,
linguistic values (Example: "low", "below average",
"average", "slow", "fast", etc.), or so-called fuzzy sets
tAAA ,...,, 21 , it is difficult to use Euclidean metric and it is
needed to introduce a fuzzy metric. In this case, another
function (membership function) which determines the
degree of fuzzy features (term set) is selected. Each term set
is described by their membership functions.
2.2 Problem Formulation Using The Membership
Function
In the process of solving some problems, the standard forms
of the membership function (bell-shaped, trapezoidal,
triangular, Gaussian type, etc.) are used. Usually these
functions are described by two or three parameters. We will
consider a membership function with two parameters (e.g.,
b, and c) [4].
Usually a specialist of the subject area determines the
appropriate types and parameters of the membership
function. In addition, builds a membership function, which
corresponds to the terms of the values of each feature based
on the input data. However, it is not always possible to
involve the experts of the subject area for solving applied
problems. Therefore, there is a need to develop the
algorithm of the settings (parameterization) of the proximity
function based on the set provided by the specialists.
A fuzzy set is denoted by a set of pairs xx
vA )ˆ( , where x
takes some informative value, and )ˆ(x
vA displays the x to
the unit interval, taking values from 0 to 1. Here )ˆ(x
vA
represents the degree of membership of x to something.
To solve the problem of parameterization, you can use
evolutionary and multi-agent methods and algorithms, such
as Genetic algorithms, neural networks, the algorithm "Ant
colony", Immune algorithm, etc. Each method and algorithm
is characterized by its own set of properties and parameters.
We consider the use of membership functions in a fuzzy
information about the features of the object, as well as
consider the problem of the parameterization of the
membership function using Genetic algorithm. By means of
this algorithm, the optimization problem is solved with the
following target (fitness) function:
),,( CBSFy 
where S - a set of objects with fuzzy values of features;
 qbbbB ...,,, 21 and  qcccC ...,,, 21 - the vectors of
parameters of the settings of the membership functions; q -
total number of terms; F - the target function, which adopts
appropriate values and determines the quality of the
optimization stage after realization of step of the ACE.
Let the set under study be in the form of m pairs of
experimental data:
  mjyS jj ,1,, 

_______________________________________________________________________________________
where  njjj
j
xxxS ...,,, 21 – input vector and jy
appropriate values of the output variable y for jth pair
<input-output>, ],[ yyyj  .
In accordance with the method of least squares the problem
of optimum settings of fuzzy model can be formulated as
follows: finding the vectors (B, C), which satisfy the
following restrictions
,,1],,[],,[ qtcccbbb tttttt 
and provide
   CB
m
j
jj yCBSF
,
2
1
min,, 

(1)
where tb and tс ( tb and tc ) lower (upper) boundary,
which gets a value of parameters of the proximity function.
3. THE SOLUTION OF THE PROBLEM
To solve the problem (1) the Genetic algorithm optimization
is proposed [4].
For the realization of the Genetic algorithm the method of
encoding the values of the unknown parameters B and C
should be defined. We reduce the parameters B and C in a
single vector:
   nn t
n
t
nnn
tt
cbcbcbcbCB ,,...,,,...,,,...,,, 11
11
1
1
1
1
11

where it - the number of terms-grade of input variable ix ,
, qttt n  ...21 ;
q – the total number of terms;
Vector  uniquely defines a model  CBSF ,, .
3.1 Crossover Operator
Since the operation of crossing is a basic operation of the
Genetic algorithm, its performance is primarily dependent
on the performance of used crossing operations. As a result,
crossing two parent chromosomes 1 and 2 provides
1Ch and 2Ch chromosome offspring through the exchange
of genes with respect to (n)th
point of the crossing.
It should be noted that because of a lot of  it
iiii aaaA ,...,, 21

terms-estimates of the input variables are in ascending order
(for example: “low”, “medium”, “high”, etc.), the
introduction of cross-breeding operation may disturb this
order. Therefore, after the exchange of genes control should
be implanted to ensure that the set of terms were ordered.
We introduce the following notation:
1
ipb – ipth
parameter of b in a chromosome-parent 1 ,
2
ipb – ipth
parameter of b in a chromosome-parent 2 ,
2Ch
ipb – ipth
parameter of b in a chromosome-parent 2Ch ,
 )1()1()1(1)1(1)1(
1
)1(
1
)1(1
1
)1(1
11 ,..,,.,,...,,..,,., 11 nn t
n
t
nnn
tt
cbcbcbcb
 )2()2()2(1)2(1)2(
1
)2(
1
)2(1
1
)2(1
12 ,..,,.,,...,,..,,., 11 nn t
n
t
nnn
tt
cbcbcbcb
After the crossover operation the following new
chromosomes are yielded:
 )2()2()1(1)1(1)2(
1
)2(
1
)1(1
1
)1(1
11 ,..,,.,,...,,..,,., 11 nn t
n
t
nnn
tt
cbcbcbcbCh 
 )1()1()2(1)2(1)1(
1
)1(
1
)2(1
1
)2(1
12 ,..,,.,,...,,..,,., 11 nn t
n
t
nnn
tt
cbcbcbcbCh 
Algorithm operation crossing two parental chromosomes 1
and 2 , which will result in offspring of 1Ch and 2Ch is as
follows:
Step 1. Generate a random number iz in an amount (n), such
that ii tz 1 , where it - the number of terms-grade input
variable nixi ,1,  .
Step 2. Genes are exchanged in accordance with the values
found i
z exchange points on the rules:







.,
,,
2
1
1
iip
iipCh
ip
zpb
zpb
b









,,
,,
1
2
2
iip
iipCh
ip
zpb
zpb
b


nitp i ,1,1  .
Step 3. Control over the order of terms is set:
    ,,   iiiiii ccbbbb 
,,1,,1 nili  
where  symbol of exchange.
The mutation operator. Each element of the vector S can
undergo operation of mutation probability mp . The
mutation of the element t is denoted by Mu(t):
    iiip xxRANDOMbMu , (2)
    iiip ccRANDOMcMu , (3)
where  ii cc , - the range of possible values of the
concentration factor-stretching features proximity terms,
estimates of input variable ix ,     ;,1,,0, nicc ii 
3.2 Mutation Operator
The algorithm of the mutation operation will be:
Step 1. For each of the element z in the vector  a
random number   1,0RANDOMz  is generated.

_______________________________________________________________________________________
If z>pm the mutation does not produce otherwise proceed to
step 2.
Step 2. Carry out the operation of the mutation element
z in accordance with formulas (2)-(3).
Step 3. Control over ordering is set.
3.3 The Function of Compliance
Denote the function of matching chromosomes  in terms
FF( ) (from the English. Fitness function). As the
compliance function we use the optimization criterion, taken
with the minus sign.
For the developed models F(S, B, C) function of ACE,
matching chromosomes derived from the criterion (1), will
be:
    


m
j
jj yCBSFFF
1
2
,, .
The minus sign is needed so that the meaning of compliance
function has not changed, that is, the lower the quality of the
proposed model, the lower the value of its compliance
function.
In accordance with the principles of Genetic algorithm
choice of parents for the operation of crossing must be
carried out not by accident. The greater the value of the
function corresponds to a certain chromosome, the greater
the likelihood that this will give the offspring chromosome.
A method for determining the parents based on the fact that
each chromosome in the population Si assigned a number
,ip such that:
    ,,1,0
1
jiji
K
i
ii ppFFFFpp  


K - number of chromosomes in the population. The number
ip is interpreted as the probability of which is calculated as
follows:
 
 

 K
j
j
i
i
FF
FF
p
1


.
Using a series of numbers ip , chromosome parents for the
operation of crossing we find the following algorithm:
Step 1. Postpone a number of ip on the horizontal axis (Fig.
2).
Figure 1. Selecting the the parent chromosomes
Step 2. We generate a random number z (Fig. 2) having a
uniform law of distribution on the interval [0,1].
Step 3. As a parent chromosome we choose i , the
corresponding sub-interval ip , which hit number z. For
example, in Fig. 2 the generated number z determines as a
parent chromosome 2 .
Step 4. We repeat steps 1-3 for the determination of the
second parent chromosome.
3.4 Generating A Population
Generating a population means determining the initial set of
solutions (chromosomes-parents) who are at the crossing.
Given initial solutions (elements of the vector S) as follows:
  iii xxRANDOMb ,0

  iii ccRANDOMc ,0

where   ,RANDOM is the operation of finding a
uniform distribution on the interval  , random number.
After the random baseline variants of chromosomes, they
should be subject to control for the order of terms are kept.
It is assumed that the initial population has a K parent-
chromosomes.
4. THE STEPS OF THE ALGORITHM
At each iteration of the Genetic algorithm population size
will increase by cpK  offspring chromosomes where cp -
the factor of crossing. To maintain a constant population
size (K), before the next iteration the worst cpK 
chromosomes (in terms of matching functions) should be
discarded. In view of this, the Genetic algorithm of the
optimal adjustment of fuzzy model F(S, B, C) ACE will be
as follows:
Step 1. We generate initial population.
Step 2. Find the values of the functions   KiFF i ,1, 
compliance by implementing the steps of the ACE.
Step 3. Define
2
cpK 
pairs of parent chromosomes.
Step 4. Perform the operation of crossing each pair of
chromosomes of parents in accordance with the algorithm
crossing.
Step 5. With probability mp perform mutation of derived
offspring chromosome according to the mutation algorithm.
Step 6. From the resulting population size of cpKK 
chromosomes cpK  chromosomes, which have the worst
value of the function of compliance  iFF  , are discarded.
Step 7. If the chromosome iS , for which   0iFF  ,
obtained, then it is the end of the algorithm, otherwise go to
step 8.

_______________________________________________________________________________________
Step 8. If the specified number of steps are not exhausted,
then proceed to step 2; Otherwise, the chromosome having
the greatest value of compliance  iFF  , received through
the ACE, is a suboptimal solution. It is the end of
algorithm.
Note that two fuzzy values considered close feature if and
only if both of the characteristic values simultaneously
belong to one term set.



 

otherwise,0
)ˆ()ˆ(,1
),(
**
iqAijA
iqij
xx
xx vv


where
)ˆ((max)ˆ(
,1
*
ijA
tv
ijA xx vv


 ,
)ˆ((max)ˆ(
,1
*
iqA
tv
iqA xx vv


 .
Here are all the range of ijx fuzzy attribute given to one
universal interval [0,] with the following ratios:
ijij
ijij
ij
xx
xx
x


ˆ
where ijij xx , - the interval of the variable ijx , ni ,1 , ijxˆ -
normalized sign.
After the calculation of proximity function )~,~( qj SSr  ,
for ~ - parts of objects jS and qS , are computed voice
),( uj KSГ to class uK , lu ,...2,1 , by 3-5 stages ACE
[1,2]:
 
 

u
u A
m
mq
qj
u
uj SSr
N
KS
1 ~
1
)~,~(
1
),(

 ,
where uq KS  , uu mmq ,11   , 1,  uuuu mmNN –
the number of objects in the class uK .
5. CONCLUSION
In conclusion, it is worth noting that the optimization of
parameters is considered as a multiextremal task. Using the
traditional mathematical methods for solving these problems
does not always produce the desired effect. Use of the
Genetic algorithm especially using the random search is
more preferred.
The considered algorithm makes it possible to build an
adaptive model ACE based on the training sample, issued by
subject matter experts.
It should be emphasized that the proposed GA, based on the
idea of a random search, may require a lot of computation
time and more resources. However, this problem can be
solved using the technology of multiprocessor parallel
computing.
REFERENCES
[1]. Zhuravlev YI, Kamilov MM, Tulyaganov Sh.E.
Algorithms for calculating estimates and their
application. -Tashkent: 1974. -124 pp. (in Russian).
[2]. Zhuravlev Y.I. Favorite scientific works. - M.: Master,
1998. - 420 pp. (in Russian).
[3]. Rothstein A.P. Intelligent identification technology,
fuzzy sets, genetic algorithms, neural networks. -
Vinnica: The UNIVERSUM, 1999. -320 p.
[4]. Rutkovskaya D., Pilinsky M., Rutkowski L. Neural
networks, genetic algorithms and fuzzy systems. M.:
Hotline - Telecom, 2006. - 452 p.
BIOGRAPHIES
He got his MSc degrees from
specialty "Informatics" in Karshi
State University 2003 respectively.
He published several papers on the
topic of Pattern Recognition, Data
mining, such as fuzzy sets theory and
the Algorithms for Calculating
Estimates in many international conferences and journals.

The solution of problem of parameterization of the proximity function in ace using genetic algorithm

Recommended

More Related Content

What's hot (18)

Viewers also liked (18)

Similar to The solution of problem of parameterization of the proximity function in ace using genetic algorithm (20)

More from eSAT Journals (20)

Recently uploaded (20)

The solution of problem of parameterization of the proximity function in ace using genetic algorithm