J3 - Elsevier JOCS - A New Population Seeding Technique For Permutation-Coded Genetic Algorithm

G Model
JOCS-208; No. of Pages 21
ARTICLE IN PRESS
Journal of Computational Science xxx (2013) xxxxxx
Contents lists available at SciVerse ScienceDirect
Journal of Computational Science

journal homepage: www.elsevier.com/locate/jocs
A new population seeding technique for permutation-coded Genetic

Algorithm: Service transfer approach
P. Victer Paul a, , A. Ramalingam b , R. Baskaran c , P. Dhavachelvan a ,
K. Vivekanandan d , R. Subramanian a
a
Department of Computer Science, Pondicherry University, Puducherry, India

Department of MCA, Sri Manakula Vinayagar Engineering College, Puducherry, India
c
Department of Computer Science and Engineering, Anna University, Chennai, India
d
Department of Computer Science and Engineering, Pondicherry Engineering College, Puducherry, India
b
a r t i c l e
i n f o
Article history:
Received 31 December 2012
Received in revised form 16 May 2013
Accepted 26 May 2013
Available online xxx
Keywords:
Genetic Algorithm
Population seeding technique
Traveling Salesman Problem
Order distance vector
Combinatorial problem
a b s t r a c t
Genetic Algorithm (GA) is a popular heuristic method for dealing complex problems with very large search
space. Among various phases of GA, the initial phase of population seeding plays an important role in
deciding the span of GA to achieve the best t w.r.t. the time. In other words, the quality of individual
solutions generated in the initial population phase plays a critical role in determining the quality of nal
optimal solution. The traditional GA with random population seeding technique is quite simple and of
course efcient to some extent; however, the population may contain poor quality individuals which
take long time to converge with optimal solution. On the other hand, the hybrid population seeding
techniques which have the benet of good quality individuals and fast convergence lacks in terms of
randomness, individual diversity and ability to converge with global optimal solution. This motivates to
design a population seeding technique with multifaceted features of randomness, individual diversity
and good quality. In this paper, an efcient Ordered Distance Vector (ODV) based population seeding
technique has been proposed for permutation-coded GA using an elitist service transfer approach. One
of the famous combinatorial hard problems of Traveling Salesman Problem (TSP) is being chosen as the
testbed and the experiments are performed on different sized benchmark TSP instances obtained from
standard TSPLIB [54]. The experimental results advocate that the proposed technique outperforms the
existing popular initialization methods in terms of convergence rate, error rate and convergence time.
2013 Elsevier B.V. All rights reserved.
1. Introduction
Genetic Algorithm (GA) is a well-known method for global optimization of complex problem very large search space based on
the survival of the ttest concept of natural evolution [55]. The
signicant features of the GA, which makes it perform competently could be dened as follows: GA operates on a population
of feasible solutions rather than on a single solution; the variety of genetic operators helps to explore unrevealed solutions in
the large search space effectively; possibility to construct problem the specic genetic operators which can offer better solution
search; population diversity helps to avoid the drawback of getting
trapped in local optima and premature convergence. These exible
Corresponding author. Tel.: +91 0413 2274430.

E-mail addresses: victerpaul@gmail.com, victerpaul.p@gmail.com
(P. Victer Paul), a.ramalingam1972@gmail.com (A. Ramalingam),
baaski@annauniv.edu (R. Baskaran), dhavachelvan@gmail.com (P. Dhavachelvan),
k.vivekanandan@pec.edu (K. Vivekanandan), subbur@yahoo.com (R. Subramanian).
congurations encourage researchers to design novel GA with a

modied operators and population seeding techniques to improve
further its performance.
GA had been proven to be efcient at searching optimal solution among a large and complex search space in an adaptable way,
controlled by the equivalent biological evolutionary mechanisms
of reproduction, crossover, and mutation. Various phases of GA
can be dened as population seeding (initial population), selection,
reproduction, crossover, mutation and termination constraint, in
which rst step occurs once and the rest of the steps are repeated
until the nal condition is satised [16,52]. The rst step of any
GA is to generate a set of possible solutions randomly as an initial
population or population seeding [34,37,41]. The quality of individual solutions in the initial population plays a critical role in
determining the quality of the nal solution that can be obtained
using GA [32,40]. However, in traditional GA, population seeding
is performed randomly which can be simple but, the whole population contains much of individuals with worst quality, infeasible
solutions sometimes [36]. As a consequence, GA with a random
population seeding technique requires longer search time to nd
1877-7503/$ see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jocs.2013.05.009
Please cite this article in press as: P. Victer Paul, et al., A new population seeding technique for permutation-coded Genetic Algorithm:
Service transfer approach, J. Comput. Sci. (2013), http://dx.doi.org/10.1016/j.jocs.2013.05.009
G Model
2
ARTICLE IN PRESS
P. Victer Paul et al. / Journal of Computational Science xxx (2013) xxxxxx
an optimal solution, number of generations required to evolve optimal solution increases, the search possibility for an optimal solution
decreases and more importantly the convergence rate or quality of
the optimal solution obtained is reduced. Thus, the requirement for
a modied population seeding technique in GA is clear and in fact,
several research works were proposed to support the dispute [61].
Lawrence and Amini [61] discussed about different GA conguration issues and claims that seeding the initial population with
heuristics can improve the efciency of the GA greatly. Togan and
Daloglu [40] believes that performance and convergence ability of
GA are critically inuenced by the population seeding method and
proposed two new self-adaptive member grouping strategies and a
new strategy for population seeding. The large search space is collected into different groups and the list of cross section values are
assigned as initial values to set the initial population automatically.
This method is dedicatedly proposed for application in the area of
structural engineering and tested in truss structures and transmission towers. Nearest Neighbor (NN) tour construction heuristic is
one of the familiar alternatives for random population seeding in
GA, particularly for TSP [3,16,4246]. In NN technique, individuals
in the population seeding are constructed with the city nearest to
the current city and such good individuals can rene the subsequent search in the next generations [3]. Though NN works ne,
it suffers with some critical factors: several cities are not included
in the individuals created initially and have to be inserted at high
costs in the end; neglecting several cities at the population seeding
stage leads to severe errors in optimal solution construction and
the diversity among the individuals created is very minimum.
Yingzi et al. [32] proposed a Greedy GA (GGA), in which the population seeding is performed using Gene Bank (GB). The GB is built
by assembling the permutation of n cities based on their distance.
In GGA method, the population of individuals is generated from the
GB such that the individuals are of above-average tness and short
dening length. In GGA, with the increase in number of cities leads
to augmented problem complexity and performance degradation,
and large collection of GB individuals enlarges the cost of computation at each generation. The improved performance of GGA is
justied using TSP with maximum of hundred cities and its performance deteriorates with large number of cities. In [38], Fuyan
et al. proposed K-means algorithm, based on the work reported in
[46], which is considered to generate much infeasible solutions, to
obtain the initial population in which N number of individuals are
partitioned and assigned to one of K clusters. As a result, using
K-means algorithm, generating infeasible solutions in the population seeding stage is avoided. Performance evaluation is performed
with a maximum of 10 cities with only time based analysis and
be decient to validate the proposed technique for large number of
cities and convergence capability. Yugay et al. [36] proposes a modied GA with sorted initial population method based on theory of
better parents produce better offsprings. In this approach, a large
initial pool of population is generated and ranked in accordance
to their tness values and at last, a certain number of individuals with bad tness are omitted. This approach also suffers with
the issues discussed with NN tour construction heuristic technique
such as premature convergence, reduced search space exploration
and minimum population diversity. Hence, the traditional GA does
not provide effective performance when applied to some of the
combinatorial problems like TSP [33], so each stage in the traditional GA has been modied in order to achieve a better output and
thus resulted in hybrid GAs [2832,34,49,50].
Recently, many researchers proposed modied versions of GA,
particularly for solving TSP using random population seeding techniques [3,16,3032,34,3941,48,6466]. Though several modied
population seeding techniques for GA have been proposed, many
researchers still continue to work with a random population seeding technique because of the complexity nature of algorithm, which
is difcult to understand and implement, problem specic modications are required to apply and problem such as premature
convergence, ineffective search space exploration and less population diversity. This implies that the researchers are interested
in random population seeding technique to accomplish a better
search space exploration and nding best optimal solution at the
cost of high convergence time. In [64], Xing et al. proposed a hybrid
approach combining an improved Genetic Algorithm and optimization strategies using random population seeding technique. An
efcient hybrid mutation Genetic Algorithm has been proposed
using a random population seeding technique in [34]. Chang and
Ramakrishna [39] proposed a GA for shortest path routing problem,
in which author reveals the reason for preferring random population seeding method than heuristic initialization. Although the
mean tness of the individuals generated using heuristic initialization are high so that it may help the GA to obtain the better
solutions faster, but it ends up in exploring a small part of the search
space and never nd the global optimal solutions because of the
lack of diversity in the initial population generated [47]. In some of
the works, the authors used hybrid population seeding technique
which combines random and any of the modied population seeding technique [33,61]. In [62], Qu and Sun proposed a synergetic
approach to GA by adding some new randomly generated individuals into the population after each generation in order to prevent
premature convergence and to obtain nal optimum.
To summarize, the modied population seeding techniques has
the advantages of good quality or generating potential sequence
individuals at an early stage and the ability to nd near optimal
solutions at few generations; however they lacks in randomness,
diversity of individuals generated, ability to explore more search
space and nding the global solution. On the other hand, the random population seeding technique has the advantages of individual
diversity, can explore search space efciently and nding optimal
solution; however they have the disadvantages of individuals with
worst potential sequence and requires longer search time to converge an optimal solution. The controversies between these two
categories of population seeding techniques motivate to propose
an efcient population seeding technique with characteristics of
randomness, individual diversity and potential sequence. Thus in
this paper, an efcient Ordered Distance Vector (ODV) based population seeding technique with three different varieties has been
proposed for permutation-coded GA. The popular combinatorial
optimization problem of Traveling Salesman Problem (TSP) is being
chosen as the testbed to validate and claim the efcacy of the proposed population seeding technique. Experiments were performed
over the different sized benchmark TSP instances obtained form
the TSPLIB [54]. The organization of the paper is as follows: Section 2 offers sufcient background information over GA and TSP to
improve the understanding of this paper. Section 3 describes the
proposed technique and its variants along with the corresponding
algorithms. The different phases of the experiments are summarized in Section 4. This section also reports experimental results
with corresponding analyses. And nally, Section 5 presents the
conclusive remarks of the work reported in this paper.
2. Background information
As stated earlier, this section offers a brief introduction on
Genetic Algorithm (GA) and Traveling Salesman Problem (TSP) to
improve the understandability of this paper.
2.1. Genetic Algorithm
Genetic Algorithms (GAs), a subclass of evolutionary algorithms,
is a stochastic optimization technique based on the principles
G Model
ARTICLE IN PRESS

Begin
Select Individual Representation
Design Fitness Function, Population Size and
Generation Limit
Generate Initial Population (Gen=0)
If
Gen>Limi
Stop
Find Best individual in Population (Bestindiv)
If Bestindiv is
Fittest
Individual
Select Individual(s)
Perform Reproduction
Perform Mutation
Perform Crossover
New Population (Gen=Gen+1) with

Population Size
Fig. 1. The generic ow of Genetic Algorithm.
and mechanisms of natural selection. In 1975, John Holland

introduced the concept of GA in his book Adaptation in Natural and Articial Systems [55]. GA is a panacea for resolving all
optimization problems and typically used to solve complex optimization problems like aircraft design, robot trajectory planning,
gas pipelining control, traveling salesman, routing, game playing,
job shop scheduling, etc. [52]. GA has two important features: it is
a stochastic algorithm where selection and reproduction are performed randomly; it always works with a population of solution,
which offers the benet of assortment and robustness to the technique. The need of GA is to nd the best solution in the large search
space, which is the collection of all feasible solutions. An individual
(or) chromosome is a representation of possible and legal solution
to a problem. A population is a collection of individuals that can be
handled by GA at a specic point of time. Fitness of an individual
is the value associated with the corresponding individual which
depends on its representation. The tness function, which is problem specic, corresponds to evaluation of quality of individual to
check for solution optimality for the problem in hand. The termination criteria for GA can be nding the best solution or reaching
maximum generation limit.
Fig. 1 illustrates a generic ow of Genetic Algorithm. GA starts
by selecting suitable format for individual representation. Followed
by designing tness function, population size and generation limit,
population size and generation limit are set as necessary. An initial
population (Gen = 0) of predened size is generated by random or
heuristics fashion. Then, the GA loops through the iteration process
till any of the termination criteria satised. Each iteration process
consists of the following steps:
Evaluate the tness value of each individual based on a tness

function.
Check whether any of the individual satises the condition for
best or optimal solution (Bestindiv ). If so, exit with the corresponding solution, otherwise continue.
Select the better individual(s) from the population to apply
genetic operators, namely reproduction, crossover and mutation, to produce next generation. These genetic operators can
be applied either individually or combination of any two. Various techniques used for selection are random [38], tournament
selection [16,52] and the roulette wheel selection [35,36].
Reproduction is copying the best (elite) solutions of the previous population to the next. The elite preservation [52] strategy
ensures the best solution of the population survives into the next
generation.
Crossover is the probabilistic process of the creating new
individuals by exchanging information of two selected parent
individuals to produce a better t individual.
Mutation is the probabilistic process to introduce constructive
changes in the information stored in the individual.
This new population generated has to replace the previous population (Gen = Gen + 1) and continue with the iterations.
Thus, at the end of GA, the optimal or near optimal solution can
be obtained for the problem considered.
2.2. Traveling Salesman Problem
The Traveling Salesman Problem (TSP) is a well-known NP-hard
problem exceedingly studied in the eld of operations research and
computer science [13]. TSP is commonly considered as a standard
testbed for various combinatorial optimization techniques. In TSP,
a salesman wants to visit each of a set of cities exactly once and
return to the starting city with minimal distance traveled. Thus
the objective of TSP is to nd a minimum total cost closed tour
that visits each city exactly once for a given number of cities and
the distance (or the cost) of traveling between any two cities. The
problem can be formalized as follows.
Let G = (Cn , En ) be the complete undirected graph such
that C {c1 , c2 , c3 , . . . , cn } and E {(c1 , c2 ), (c1 , c3 ), . . . , (c1 , cn ), (c2 ,
c3 ), (c2 , c4 ), . . . , (c2 , cn ), . . . , (c(n1) , cn )}. In the graph G, C and E
corresponds to the cities and path between the cities respectively.
For each pair, (ci , cj ) and i =
/ j, Distance between the city ci and
cj can be given as d(ci , cj ) TSP aims to nd the minimum distance
tour (optimal solution) between cities which can be given as,
Optimal solution = min
n
d(ci , c(i+1) ),
1 (n + 1)
(1)
i=1
The search space for the TSP is a set of permutation of n cities.

Any permutation of n cities gives a possible solution and the size of
the TSP search space with n cities is ((n 1)!/2). Let be the set of
permutation of cities (search space),
= {1 , 2 , . . . , n! }
(2)
and
1...n! = {ci , c(i+1) , c(i+2) , . . . , cn }
(3)
TSP has to search for the optimal (best) solution which exists
in the permutation set and can be given as,
n
Check whether number of generation exceeds the generation

limit (Gen > limit). If so, exit with best solution of previous generation, otherwise continue.
i=1
d(c i , c (i+1) )
n
d(c i , c (i+1) )
(4)
i=1
Such that, 1 (n + 1) and {}
G Model
ARTICLE IN PRESS
Thus, be the required best solution for the considered TSP with
n number of cities.
By considering the importance and extended applications of
TSP in various elds, several meta-heuristic search techniques have
been proposed to solve it, such as Tabu Search [9,10], Genetic Algorithm [1117], Ant Colony Optimization [1822], Particle swarm
Optimization [23,29], Neural Networks [24,25], Simulated Annealing [26], Multi-Agent System [27,35,41,63] and Hybrid-Heuristics
[2830]. Major applications of TSP are Vehicle routing [4], Drilling
of printed circuit boards [5], Overhauling gas turbine engines [6],
X-ray crystallography [7], computer wiring [4], sequencing and
scheduling [47,52] and the order-picking problem in warehouses
[8]. In this paper, TSP is being used as a testbed for demonstrating
the effectiveness of the proposed technique. On the other hand, it
could also be considered that the proposed technique is a better
solution for solving TSP.
3.1. Problem statement

As described in Section 1, the controversies between the random
and heuristic population seeding techniques motivated to propose
an efcient population seeding technique with the characteristics
of randomness, individual diversity and potential sequence. The
proposed Ordered Distance Vector (ODV) based Population Seeding, which depends on the ODV matrix, is different from those in
the literature in terms of its simplicity, fastness and global optimal
solution convergence. The features such as randomness, potential
sequence and diversity are considered as critical factors to improve
the effectiveness of the proposed ODV population seeding technique.
3.2. Ordered Distance Vector matrix
For any TSP, ODV for any city ci can be formed by the permutation of (n 1) cities, which are arranged in the increasing order of
distance from the city ci . ODV for the city ci can be given as,
ODV (ci ) = (cj , cj+1 , cj+2 , . . . , cn1 )
(5)
such that,
d(ci , cj ) d(ci , cj+1 ) d(ci , cj+2 ) . . . d(ci , cn1 )
(6)
where, d(ci , cj ) refers to the distance between the cities ci and cj .

An ODV Matrix is a (n (n 1)) matrix, which contains OD for
each city and arranged in the sequence of its distance matrix D. In
general ODM can be represented as,
ODVc1
C1(j)

ODVc2 C2(j)

ODM =
ODVc3 = C3(j)

... ...
ODVcn
Ordered Distance Vector based population seeding technique

generates a set of permutation of n cities using the Ordered Distance Vector matrix. In each permutation, the sequence of the cities
is chosen such that the sum of distances between the cities is near
minimum. The initial population generated using ODV population
seeding technique PODM can be represented as,
PODM
1 (ci )
1 (c(i+1) )
1 (c(i+2) ) . . .
1 (cn )
2 (ci ) 2 (c(i+1) ) 2 (c(i+2) ) . . . 2 (cn )
=
3 (ci ) 3 (c(i+1) ) 3 (c(i+2) ) . . . 3 (cn )
...
...
...
...
o (ci )
o (c(i+1) )
o (c(i+2) )
(10)
o (cn )
where k is the kth permutation of n cities, k = (1, 2, 3, . . ., o). o

is the total number of individuals generated in the population.
There are two signicant characteristics that makes the ODV
population seeding technique distinct from the others and they are
as follows:
3. Proposed system
3.3. ODV based population seeding techniques
Cn(j)
C1(j+1)
C1(j+2)
C1(n1)
C2(j+1)
C2(j+2)
C2(n1)
C3(j+1)
C3(j+2)
...
...
Cn (j + 1)
Cn(j+2)
C3(n1)
...
(7)
Cn(n1)
Rank of distance of city cx with cy denoted as R(cx , cy ) refers to

the position of the city cx in the row corresponding to the city cy in
the ODM.
R(d(c1(j+2) , c1 )) = 3
(8)
R(d(c2(n1) , c2 )) = n 1
(9)
The rank of distance of the city is used to choose the city at

specic position in the ODM matrix for solution generation.
Potential sequence: This factor helps to keep the good quality order
of cities in each individual.
Individual diversity: This factor refers to the difference in the permutation of cities in each individual generated. It helps to avoid
the drawback of getting trapped in local optima and premature
convergence.
Potential sequence: For GA, the capability to nd the optimal
solution is critically affected by the tness (potential sequence) of
individuals generated in the population initialization [9]. This factor
can be applied in ways as follows:
Equi-begin (Eb): It is claimed that forcing individual to always
begin the permutation of city at a specic city cx will allow the GA
to take the benet of good building block or high tness sequence
of cities in generating new individuals [7,10].
Vari-begin (Vb): This method is used in many existing techniques
in which the beginning of the permutation of the city is not necessarily the same city. Creating the individuals beginning with
different cities provide the possibility of creating new individuals, the size of the initial population can be more, repetition of
individual sequence can be avoided and mainly it helps to avoid
the algorithm to get trapped in premature convergence.
Individual diversity: Individual diversity refers to the variation
in the permutation of cities in each individual in the population.
The initial population of the individual must offer a wide diversity because preserving diversity in the population, particularly
at the initial stages of GA, represents a condition for avoiding the
search from getting stuck in local optima and also from premature
convergence.
Best adjacent (ba) number: A detailed study on existing work
[5658] helps to derive an assumption that, in an optimal solution,
any city ci is connected to city cj such that cj is one of the ci s nearest
ba number of cities. The value of ba (an integer) can be expressed
in range and it highly depends on the size of the population n. The
range of ba value can be assigned as follows:
If 1 n 10, then 2 ba 3.
If 11 n 100, then 2 ba 4.
If 101 n 1000, then 2 ba 5.
If n 1001, then 2 ba 6.
G Model
ARTICLE IN PRESS
technique as shown in Fig. 2 and the comparison among those variants are given in Table 1. These variants are explained as follows:
3.3.1. Type 1: Equi-begin with Variable diversity (EV)

In this type, a new bax number is generated before adding each
city into every individual and each individual starts with the same
city. The individuals created in this type have high potential permutation of cities and thus the convergence time can be reduced.
The maximum number of individuals in the initial population that
is being created using this type can be given as:
Fig. 2. Types of ODV based population seeding techniques.
Max(tot(PODM )) = ba(n1)
(11)
The above said ranges are not exact and are derived as the
experimental outcomes of the literature study. It shows that if the
number of cities n is between 101 n 1000, then in the optimal
solution, every city would be connected to any of its ba 2 ba 5
cities which are nearest to it.
In ODV population seeding technique, the value of ba number plays a vital role in providing individual diversity. The adjacent
cities in each individual of the initial population depends on the
ba value. An integer value generated between the ranges of ba
decides which cities are kept as adjacent in each individual. There
are two different ways of initializing population based on diversity
in ba number generation,
where tot(PODM ) refers to the total number individuals in the population PODM and ba refers to the Best Adjacent number assigned
and n refers to the total number of cities.
The PODM generated using this type can be represented as:
Equal ba diversity (Ed): An integer value bax is generated

between the ranges of ba before creating each individual, for
every city, the city at bax position is added as adjacent. The rank
of distance between each city in the same individual is same. If it
is already added then the city at next position is added and the
same is repeated until the complete individual is created. In this
method, bax value is generated once and the diversity (rank of
distance) between each city is equal for each individual in the
population.
Variable ba diversity (Vd): An integer value bax is generated
between the ranges of ba before adding each city in the individual, for every city, the city at currently generated bax position is
added as adjacent city. The rank of distance between each city in the
same individual is different. In this method, bax value is generated
for n 1 times and the diversity between each city is different
for each individual.
where the rst city remains same for each individual (i.e.) 1 (1)
2 (1) 3 (1), . . . , o (1).
This type of population should be handled with suitable
crossover operators because it may be suffered by pre-convergence
problem due to lack of much diversity.
Based on these two characteristic factors, there are three

effective ways for generating initial population in ODV seeding
PODM
1 (1)
1 (c2 )
1 (c3 ) . . .
1 (cn )
2 (1) 2 (c2 ) 2 (c3 ) . . . 2 (cn )
=
3 (1) 3 (c2 ) 3 (c3 ) . . . 3 (cn )
...
...
...
...
o (1)
o (c2 )
o (c3 )
(12)
o (cn )
3.3.2. Type 2: Vari-begin with Equal diversity (VE)

In this type, a new bax number is generated for each individual
created and the same bax number is used to add every the cities in
individual. In the population, there is exactly ba number of individuals that starts with same initial city. The maximum number
of individuals in the initial population that is be created using this
type can be given as,
Max(tot(PODM )) = n ba
(13)
where tot(PODM ) refers to the total number individual in the population PODM and ba refers to the Best Adjacent number assigned
Table 1
Comparison between the variants of ODV population seeding technique.
Population seeding techniques
Characteristics
Initial city of individuals
Individual diversity
Rank of distance of cities

No. of times bax generated
Max. possible initial population
Advantages
Disadvantages
Type 1 (EV)
Type 2 (VE)
Type 3 (VV)
Same
Random bax value
Huge diversity among
individuals
Different
(n 1) o
ba(n1)
High potential sequence of
cities
Limited size of population
Different
Same bax value
Little diversity among
individuals
Equal
n ba
n ba
Limited size of population
Suitable crossover operators

have to used to avoid
pre-convergence problem
Suffer cyclic repetition

problem
Stuck in local optimal
solution
Different
Random bax value
Huge diversity among
individuals
Different
no
ban
High potential sequence of
cities
No cyclic repetition problem
No pre-convergence problem
Very large size of population
G Model
ARTICLE IN PRESS
Fig. 3. List of variables used in the algorithm for ODV based population seeding technique.
The PODM generated using this type can be represented as:
PODM
1 (1)
(1)
2
.
.
.
ba (1)
(2)
1
=
.
..
(2)
ba
1 (n)
.
..
ba (n)
1 (c2 )
1 (c3 )
2 (c2 )
2 (c3 )
..
.
..
.
ba (c2 )
ba (c3 )
2 (c2 )
3 (c3 )
..
.
..
.
ba (c2 )
ba (c3 )
2 (c2 )
2 (c3 )
..
.
..
.
ba (c2 )
ba (c3 )
1 (cn )
solution with minimum convergence time. The maximum number

of individuals in the initial population that is be created using this
type can be given as,
..
ba (cn )
3 (cn )
..
ba (cn )
2 (cn )
..
.
2 (cn )
Max(tot(PODM )) = ban
(15)
where tot(PODM ) refers to the total number individual in the population PODM and ba refers to the Best Adjacent number assigned
The PODM generated using this type can be represented as,
(14)
ba (cn )
where The rst city remains same for ba number of individuals

(i.e.)
1 (1) 2 (1) . . . ba (1), 1 (2) 2 (2)
. . . ba (2) and so on and R(1 (c2 ), 2 (1)) R(1 (c3 ), 1 (c2 ))
. . . R(1 (cn ), 1 (cn1 ))
This type of population seeding is not so efcient because the
individuals created using this technique suffer from cyclic repetition problem and only ba number of individuals are unique.
The cyclic repetition problem could be referred as the sequence of
the adjacent cities remains identical for more than one individual
though its initial and nal cities are different. The crossover and
mutation rate has to be set high in order to overcome the cyclic
repetition problem.
3.3.3. Type 3: Vari-begin with Variable diversity (VV)
In this type, a new bax number is generated before adding each
city into every individual and each individual starts with the random city. The individuals created in this type have high potential
permutation of cities and also have good individual diversity. This
type of population seeding is most effective and can produce best
PODM
(c ) (c ) (c ) . . . (c )
1 1
1 2
1 3
1 n
2 (c1 ) 2 (c2 ) 2 (c3 ) . . . 2 (cn )
= 3 (c1 ) 3 (c2 ) 3 (c3 ) . . . 3 (cn )
...
o (c1 )
...
o (c2 )
...
o (c3 )
...
o (cn )
This type of population seeding technique is the most recommended for effective search for best solution. The main difculty
concerned with this technique is to determine the limit of population size o, to support small population size techniques, because
good initial population individuals which may occur outside the
limit are left unrevealed.
As the last, the fourth possible type population seeding technique, Equi-begin with Equal diversity (EE), which is not considered
as an effective one since it can produce the initial population with
only n individuals. It is very hard to nd the optimal solution with
less initial population containing un-potential sequence of cities,
which leads to pre-mature convergence and large number of generations.
3.4. Algorithm for ODV population seeding technique
The algorithm for ODV population seeding technique consists
of two stages. First an Order Distance Matrix (ODM) will be created from the Distance Matrix (DM) and followed by generating
the initial population based on the ODM. The second stage of the
technique can be performed in three ways; EV, VE and VV depends
on the type of population seeding technique to be chosen. List of
G Model
ARTICLE IN PRESS
Fig. 4. Algorithm for ODV based population seeding technique and its variants.
variables used in the proposed algorithm is shown in Fig. 3. The

proposed ODV population seeding algorithm, as shown in Fig. 4,
consists of ve procedures namely Generate INIT POP(), Generate ODM(DM), EV(ODM), VE(ODM), and VV(ODM). The algorithm
begins with the procedure Generate INIT POP() and returns the
initial population. This procedure gets n and DM as inputs and

invoke other procedures conditionally. The functionalities of other
procedures are explained as follows:
Generate ODM (DM): This procedure creates ODM from the
DM, which is passed as an argument to the procedure. In this
G Model
ARTICLE IN PRESS
Table 2
Experimental results of random population initialization method.
S. no. Instance Optimal
tness
1
2
3
4
5
6
7
Eil51
Pr76
KroA100
Pr144
Gil262
Fl417
Pr1002
426
108,159
21,282
58,537
2378
11,861
259,045
Time taken
(s)
0.51
0.61
1.25
1.35
3.51
7.12
15.04
Best tness Worst

tness
803.56
195,678.73
39,789.3
111,987.32
4469.94
22,646.99
511,527
1176.34
247,890.85
58,475.04
154,899.24
6402.06
33,757.7
755,689.63
Average
tness
931.38
21,7356.19
44,782.09
124,689.52
5173.77
27,848.63
627,364.63
procedure, each row from the DM is retrieved and sorted, the

changes in each of DM is tracked with the corresponding changes
in the sequence of cities in the row. The sequence of cities
with corresponding changes, neglecting the rst element, which
will be always zero (symmetric TSP), is assigned to a row in
ODM. The sorting technique followed in the procedure is Bubble
Sort.
EV (ODM): This procedure will be invoked if the type of population seeding is Type 1, which takes ODM as argument and returns
the initial population (Pop) generated using equi-begin with the
variable diversity (EV) method. This procedure gets initial city and
ba value either from the user or may be a random value chosen
by the system (ba range will be chosen based on problem size).
Each individual is generated using EV method until the number of
individuals in generating population reaches the insisted Pop Size
value.
VE (ODM): This procedure will be invoked if the type of population seeding is Type 2, which takes ODM as argument and returns
the initial population generated using Vari-begin with Equal diversity (VE) method. For each individual, the initial city has been
selected randomly and the value of ba is chosen by the system
based on the problem size. In this procedure, each individual is
Error rate
Convergence
rate
Best indiv.
(%)
Worst
indiv. (%)
Best indiv.
(%)
Worst
indiv. (%)
88.63
80.92
86.96
91.31
87.97
90.94
97.47
176.14
129.19
174.76
164.62
169.22
184.61
191.72
11.37
19.08
13.04
8.69
12.03
9.06
2.53
76.14
29.19
74.76
64.62
69.22
84.61
91.72
Convergence
diversity (%)
Average
convergence
(%)
87.51
48.27
87.80
73.31
81.25
93.67
94.25
18.63
0.96
10.42
13.01
17.57
34.79
42.18
generated using VE method and returns the generated population

(Pop) until the Pop Size limit.
VV (ODM): This procedure will be invoked if the type of population seeding is Type 3, which takes ODM as argument and returns
the initial population (Pop) generated using Vari-begin with Vari
diversity (VV) method. In this procedure, the value of the initial
city and the ba value have been chosen randomly for the addition
of each individual and city in the individual respectively. Once the
population size reaches the Pop Size limit, the procedure stops and
returns the generated population.
At the end of the algorithm, an initial population of size Pop Size
has been generated using one of these three population seeding
methods and stored in the variable POP.
4. Experimentation and result analysis
4.1. Experimentation setup essentials
4.1.1. Individual representation
The signicance of the proposed technique is demonstrated
by assessing their performance w.r.t. the problem space of
Traveling Salesman Problem (TSP). Holland [55] classied the
Table 3
Experimental results of Nearest Neighbor population initialization method.
tness
1
2
3
4
5
6
7
Eil51
Pr76
KroA100
Pr144
Gil262
Fl417
Pr1002
426
108,159
21,282
58,537
2378
11,861
259,045
Time taken
(s)
1.18
1.26
1.96
2.01
5.62
12.46
20.12
Best tness
495.77
121,567.43
24,421.8
62,952.97
2622.75
14,123.99
312,135.76
Worst
tness
688.01
186,540.75
32,689.65
82,577.56
3532.06
18,554.38
409,843.75
Average
tness
572.65
169,746.4
29,765.76
72,458.43
2734.75
17,754.67
379,744.72
Error rate
Convergence
rate
Best indiv.
(%)
Worst
indiv. (%)
Best indiv.
(%)
Worst
indiv. (%)
16.38
12.40
14.75
7.54
10.29
19.08
20.49
61.50
72.47
53.60
41.07
48.53
56.43
58.21
83.62
87.60
85.25
92.46
89.71
80.92
79.51
38.50
27.53
46.40
58.93
51.47
43.57
41.79
Convergence
diversity (%)
Average
convergence
(%)
45.13
60.07
38.85
33.53
38.24
37.35
37.72
65.58
43.06
60.14
76.22
85.00
50.31
53.41
Convergence
diversity (%)
Average
convergence
(%)
48.25
29.30
51.28
36.45
37.31
39.81
35.29
60.58
52.59
60.66
72.96
68.03
45.32
47.70
Table 4
Experimental results of Gene Bank population Initialization method.
tness
1
2
3
4
5
6
7
Eil51
Pr76
KroA100
Pr144
Gil262
Fl417
Pr1002
426
108,159
21,282
58,537
2378
11,861
259,045
Time taken
(s)
1.14
1.03
1.96
2.18
4.56
9.96
16.62
Best tness
505.65
130,237.83
25,059.4
69,189.73
2937.24
15,190.99
340,468.3
Worst
tness
711.21
161,930.85
35,973.04
90,528.27
3824.45
19,913.1
431,877.73
Average
tness
593.91
159,436.19
29,654.83
74,363.72
3138.27
18,346.11
394,532.29
Error rate
Convergence
rate
Best indiv.
(%)
Worst
indiv. (%)
Best indiv.
(%)
Worst
indiv. (%)
18.70
20.41
17.75
18.20
23.52
28.08
31.43
66.95
49.72
69.03
54.65
60.83
67.89
66.72
81.30
79.59
82.25
81.80
76.48
71.92
68.57
33.05
50.28
30.97
45.35
39.17
32.11
33.28
G Model
S. no.
Instance
Optimal
tness
ba value
Time taken
(s)
Best tness
Worst
tness
Average
tness
Error rate
Best indiv.
(%)
Convergence
rate
Worst
indiv. (%)
Best indiv.
(%)
Worst
indiv. (%)
Convergence
diversity (%)
Average
convergence
(%)
Eil51
426
2
3
4
5
1.04
1.13
1.13
1.16
480.23
468.72
459.91
486.02
648.01
675.38
660.55
674.68
552.38
555.20
565.52
586.51
12.73
10.03
7.96
14.09
52.11
58.54
55.06
58.38
87.27
89.97
92.04
85.91
47.89
41.46
44.94
41.62
39.38
48.51
47.10
44.29
70.33
69.67
67.25
62.32
Pr76
108,159
2
3
4
5
0.89
1.04
0.96
0.98
122,815.73
129,988.13
122,222.99
117,737.35
161,930.85
180,861.62
181,133.91
185,199.42
150,166.19
152,180.65
156,608.47
160,039.16
13.55
20.18
13.00
8.86
49.72
67.22
67.47
71.23
86.45
79.82
87.00
91.14
50.28
32.78
32.53
28.77
36.16
47.04
54.47
62.37
61.16
59.30
55.21
52.03
KroA100
21,282
2
3
4
5
1.83
2.36
1.96
2.57
22,579.30
23,835.01
24,705.75
27,226.07
31,671.04
33,867.96
36,461.39
36,658.67
26,797.09
28,967.86
30,623.82
31,916.62
6.10
12.00
16.09
27.93
48.82
59.14
71.33
72.25
93.90
88.00
83.91
72.07
51.18
40.86
28.67
27.75
42.72
47.14
55.24
44.32
74.09
63.89
56.10
50.03
Pr144
58,537
2
3
4
5
1.97
1.89
1.96
2.47
62,952.97
65,118.87
67,942.88
68,776.48
77,253.26
84,841.73
93,087.71
104,878.52
68,031.87
71,417.70
78,660.40
81,943.07
7.54
11.24
16.07
17.49
31.97
44.94
59.02
79.17
92.46
88.76
83.93
82.51
68.03
55.06
40.98
20.83
24.43
33.69
42.96
61.67
83.78
78.00
65.62
60.01
Gil262
2378
2
3
4
5
4.61
4.56
4.94
4.50
2695.27
2775.68
3039.58
3094.87
3419.06
3608.30
3797.79
3860.18
3138.27
3293.49
3395.44
3504.37
13.34
16.72
27.82
30.15
43.78
51.74
59.71
62.33
86.66
83.28
72.18
69.85
56.22
48.26
40.29
37.67
30.44
35.01
31.88
32.18
68.03
61.50
57.21
52.63
Fl417
11,861
2
3
4
5
9.64
9.83
9.98
9.85
13,320.99
13,742.98
14,738.10
15,034.46
18,086.68
18,992.64
19,650.10
20,188.81
16,149.93
16,856.82
17,295.57
17,858.83
12.31
15.87
24.26
26.76
52.49
60.13
65.67
70.21
87.69
84.13
75.74
73.24
47.51
39.87
34.33
29.79
40.18
44.26
41.41
43.46
63.84
57.88
54.18
49.43
Pr1002
259,045
2
3
4
5
15.95
20.43
24.53
19.97
311,056.00
318,419.76
344,430.94
360,853.18
362,563.17
381,566.91
392,676.91
406,318.12
335,898.99
338,586.44
367,596.27
380,040.78
20.08
22.92
32.96
39.30
39.96
47.30
51.59
56.85
79.92
77.08
67.04
60.70
60.04
52.70
48.41
43.15
19.88
24.38
18.62
17.55
70.33
69.29
58.10
53.29
ARTICLE IN PRESS
Table 5
Experimental results of EV population seeding method.
G Model
ARTICLE IN PRESS
10
Fig. 5. The range of variation in convergence rate of EV method.
Fig. 6. The range of variation in error rate of EV method.
individual representation of tour of cities in the TSP into Path, Ordinal, Position Listing adjacency and adjacency list representation
schemes. Among these schemes, the Path representation of an individual is a most widely used scheme for the TSP [12,35], which is
considered for the experiments reported in this paper.
4.1.2. Performance factors
There are four performance factors used to investigate the signicance of the proposed technique and they are summarized as
follows:
(i) Convergence rate: Convergence rate of a solution can be dened
as the percentage of tness attained by the solution w.r.t. the
known optimal solution for the problem. It can be given as,
Convergence rate(%)=1
FitnessOptimal tness
100 (17)
Optimal tness
(ii) Average convergence: It is the average of convergence rate of

solutions in the given population. It can be given as,
Average convergence(%)=1
Average tness Optimal tness

Optimal tness
100
(18)
where Average tness is the average tness value of solutions in

the population and Optimal tness is the known optimal value
of the corresponding instance.
(iii) Error rate: Error rate of a solution can be dened as the percentage of difference in the tness value of the solution with
the known optimal solution for the problem. It can be given as
Error rate(%) =
Fitness Optimal tness

100
Optimal tness
(19)
(iv) Convergence diversity: The convergence diversity of the population is the difference between the convergence rate of the
best and worst solutions found in the population. This factor
shows the diversity among the individuals in the population,
which is an important factor used to overcome the pre-mature
convergence problem and it can be represented as,
Convergence diversity(%) = CRBest Indiv CRWorst Indiv
(20)
where CRBest Indiv refers to the convergence rate of best solution

in the population and CRWorst Indiv refers to the convergence rate
of worst solution in the population.
4.1.3. Experimental phases

All the implementations are done using MATLAB with TSP
benchmark datasets obtained from TSPLIB [54]. The TSP instances
that have been chosen for experimentation are Eil51, Pr76,

KroA100, Pr144, Gil262, Fl417 and Pr1002. The experiments in this
research are carried out in two different phases:
Phase I: The scope of the experiments in this phase is limited to
the initialization phase alone and not covers the full life cycle of
the Genetic Algorithm. This restricted scope of the experiments
help to assess the performance of the proposed technique in its
intended phase alone.
Phase II: The scope of the experiments in this phase is extended to
cover the whole life cycle of the Genetic Algorithm. This extended
scope of the experiments facilitates to assess the overall impact
of the proposed technique as a whole.
In both the phases, the performance of the proposed technique
has been investigated and compared with the existing techniques
like *Random Initialization technique, Nearest Neighbor technique
and Gene Bank technique.
4.2. Phase I: Experimentation and result analyses
In this phase, the quality of individuals generated by the proposed and the existing population seeding techniques have been
assessed and evaluated. The size of population for initialization is
set as 100. For each technique, experimental outcomes have been
collected for 50 runs and the average values of 50 runs are used for
analyses.
4.2.1. Experiments on existing techniques
4.2.1.1. Random Initialization technique. In random population initialization, the individuals are generated by choosing random
adjacent cities. A uniform random number generator has been used
to improve the search space exploration. Table 2 depicts the performance of the population generated for different TSP instances
using Random Initialization technique. From Table 2, the following
observations can be made:
Observation 1: For each problem instance, the time taken for
generation of initial population is minimum for random technique.
Observation 2: Random population initialization offers maximum convergence diversity since uniform random number
generator has been used. Consequently, the convergence rate of
best individual and the average convergence rate of the population are not convincingly potential enough (<0%) to converge to an
optimal solution.
4.2.1.2. Nearest Neighbor (NN) technique. In Nearest Neighbor population seeding technique, the city y can be selected as adjacent
city for the city x such that it would be the nearest unvisited city
G Model
S. no.
Instance
Optimal
tness
ba value
Time taken
(s)
Best tness
Worst
tness
Average
tness
Error rate
Convergence
rate
Best indiv.
(%)
Worst
indiv. (%)
Best indiv.
(%)
Worst
indiv. (%)
Convergence
diversity (%)
Average
convergence
(%)
Eil51
426
2
3
4
5
0.85
0.68
0.79
0.64
514.61
510.67
510.67
510.67
635.02
635.02
638.16
688.21
571.92
568.51
570.44
583.60
20.80
19.88
19.88
19.88
49.07
49.07
49.80
61.55
79.20
80.12
80.12
80.12
50.93
50.93
50.20
38.45
28.26
29.19
29.93
41.68
65.75
66.55
66.09
63.00
Pr76
108,159
2
3
4
5
0.58
0.71
0.66
0.67
142,396.36
142,396.36
142,396.36
142,396.36
159,821.61
170,453.18
187,050.56
187,050.56
152,344.17
155,229.69
161,093.74
161,892.75
31.65
31.65
31.65
31.65
47.77
57.60
72.94
72.94
68.35
68.35
68.35
68.35
52.23
42.40
27.06
27.06
16.11
25.94
41.29
41.29
59.15
56.48
51.06
50.32
KroA100
21,282
2
3
4
5
1.50
1.36
1.54
1.59
25,125.22
25,125.22
25,125.22
25,125.22
28,541.55
31,983.78
36,397.00
37,333.38
26,544.47
27,817.70
30,195.35
31,326.19
18.06
18.06
18.06
18.06
34.11
50.29
71.02
75.42
81.94
81.94
81.94
81.94
65.89
49.71
28.98
24.58
16.05
32.23
52.96
57.36
75.27
69.29
58.12
52.80
Pr144
58,537
2
3
4
5
1.31
1.50
1.48
1.39
61,496.32
61,496.32
61,496.32
61,496.32
80,851.07
80,851.07
84,518.93
97,687.98
68,600.74
71,180.47
72,375.21
72,375.21
5.06
5.06
5.06
5.06
38.12
38.12
44.39
66.88
94.94
94.94
94.94
94.94
61.88
61.88
55.61
33.12
33.06
33.06
39.33
61.83
82.81
78.40
76.36
76.36
Gil262
2378
2
3
4
5
3.87
3.85
4.34
4.51
3002.62
3002.62
3002.62
3002.62
3593.52
3685.80
3973.25
4120.66
3199.15
3343.51
3407.15
3487.54
26.27
26.27
26.27
26.27
51.12
55.00
67.08
73.28
73.73
73.73
73.73
73.73
48.88
45.00
32.92
26.72
24.85
28.73
40.82
47.02
65.47
59.40
56.72
53.34
Fl417
11,861
2
3
4
5
8.94
7.96
8.23
8.11
15,104.49
15,104.49
15,104.49
15,104.49
19,043.95
19,043.95
19,043.95
19,607.30
16,820.81
17,326.59
17,628.11
17,492.75
27.35
27.35
27.35
27.35
60.56
60.56
60.56
65.31
72.65
72.65
72.65
72.65
39.44
39.44
39.44
34.69
33.21
33.21
33.21
37.96
58.18
53.92
51.38
52.52
Pr1002
259,045
2
3
4
5
17.40
17.10
17.11
18.90
315,597.59
315,597.59
315,597.59
315,597.59
351,116.55
377,213.31
410,111.38
436,384.90
329,395.18
345,410.40
360,368.97
374,376.88
21.83
21.83
21.83
21.83
35.54
45.62
58.32
68.46
78.17
78.17
78.17
78.17
64.46
54.38
41.68
31.54
13.71
23.79
36.49
46.63
72.84
66.66
60.89
55.48
ARTICLE IN PRESS

11
Table 6
Experimental results of VE population seeding method.
G Model
12
ARTICLE IN PRESS
Fig. 7. The range of variation in convergence rate of VE method.
of the city x. Experimental results of NN population seeding technique are shown in Table 3. From Table 3, the following observation
can be made:
Observation 3: For all the problem instances, NN technique
produces the individual with least error rate, consequently the best
individual, when compared to the other existing techniques.
4.2.1.3. Gene Bank (GB) technique. In this technique, the gene bank
is built by assembling the permutation of N cities based on their
distance. The population of individuals is generated from the gene
bank such that the individuals are of above-average tness and
short dening length. Table 4 shows the experimental results of GB
population seeding technique and from that the following observation can be made:
Observation 4: It can be observed that the quality of best individuals generated decreases with increase in the size of problem
instance. It is because of the constant gene size used for the generation of initial population.
4.2.2. Experiments on proposed techniques
4.2.2.1. Type 1: EV technique. The experimental results of EV
method of ODV based population seeding technique is presented in
Table 5. In Table 5, the best results are emphasized in bold. The following observations can be made based on the statistical analyses
shown in Table 5.
Observation 5: For all the problem instances and sizes, the EV
population seeding technique tends to give the best t individual
with the ba value of 2.
Observation 6: The error rate of the best individuals generated using the EV population seeding technique increases with an
increase in the number of cities in the instance.
Fig. 8. The range of variation in error rate of VE method.
Fig. 9. The range of variation in convergence rate of VV method.
Observation 7: The Convergence rate of the best individuals created using the EV population seeding technique decreases with an
increase in the number of cities in the instance. The maximum and
minimum values obtained are 93.90% for instance KroA100 with
ba value of 2 and 80.92% for instance Pr1002 with ba value 2.
This is because with the same population size of 100, the initial
convergence rate of the technique decreases with an increase in the
number of cities in the instance. However, at least 80.92% of convergence has been achieved at the population initialization stage
of the GA for solving TSP problem.
Observation 8: The convergence diversity of the individuals
remains constant for all the instances except Pr1002 for which
it decreases comparatively. The most suitable practical diversity
range can be 3545%. The maximum and minimum values obtained
are 62.37% for instance of Pr76 with ba value of 2 and 17.55% for
Pr1002 with ba value of 5.
Observation 9: The average convergence of EV technique
remains between the range of 6050%, which shows that the
population contains a good composition of good and bad quality
individuals leads for exploring more search space.
The ranges of variations in convergence rate and error rate of the
population generated by EV technique are illustrated in Figs. 5 and 6
respectively.
4.2.2.2. Type 2: VE technique. The experimental results of the VE
method of ODV based population seeding technique is recorded in
Table 6 and the best results are emphasized in bold. The following
observations are made based on the results presented in Table 6.
Fig. 10. The range of variation in error rate of VV method.
G Model
S. no.
Instance
Optimal
tness
ba value
Time taken
(s)
Best tness
Worst
tness
Average
tness
Error rate
Best indiv.
(%)
Convergence
rate
Worst
indiv. (%)
Best indiv.
(%)
Worst
indiv. (%)
Convergence
diversity (%)
Average
convergence
(%)
Eil51
426
2
3
4
5
1.16
0.96
1.12
1.30
481.93
466.72
464.15
488.30
650.60
681.36
677.47
694.68
562.88
556.12
572.71
589.75
13.13
9.56
8.96
14.63
52.72
59.94
59.03
63.07
86.87
90.44
91.04
85.37
47.28
40.06
40.97
36.93
39.59
50.38
50.08
48.45
67.87
69.46
65.56
61.56
Pr76
108,159
2
3
4
5
0.96
1.10
1.24
0.85
132,815.73
116,449.76
121,277.86
133,252.43
164,233.18
172,650.39
184,779.43
186,572.67
151,647.85
153,976.41
157,896.94
162,671.50
22.80
7.67
12.13
23.20
51.84
59.63
70.84
72.50
77.20
92.33
87.87
76.80
48.16
40.37
29.16
27.50
29.05
51.96
58.71
49.30
59.79
57.64
54.01
49.60
KroA100
21,282
2
3
4
5
2.09
2.29
2.33
2.03
23,081.90
24,763.78
25,720.73
27,389.15
32,911.35
34,533.24
37,224.15
37,511.53
27,110.27
29,062.30
30,894.57
31,862.64
8.46
16.36
20.86
28.70
54.64
62.27
74.91
76.26
91.54
83.64
79.14
71.30
45.36
37.73
25.09
23.74
46.19
45.90
54.05
47.56
72.61
63.44
54.83
50.28
Pr144
58,537
2
3
4
5
1.81
2.23
1.83
1.91
63,213.76
65,249.00
67,995.22
68,069.09
78,883.82
82,901.59
92,248.32
107,500.38
68,178.38
72,129.39
79,161.49
78,761.06
7.99
11.47
16.16
16.28
34.76
41.62
57.59
83.65
92.01
88.53
83.84
83.72
65.24
58.38
42.41
16.35
26.77
30.16
41.43
67.36
83.53
76.78
64.77
65.45
Gil262
2378
2
3
4
5
5.04
5.29
4.85
5.14
2639.70
2728.50
3054.15
3145.82
3443.90
3731.19
3772.09
3986.36
3125.53
3294.28
3407.21
3523.60
11.01
14.74
28.43
32.29
44.82
56.90
58.62
67.63
88.99
85.26
71.57
67.71
55.18
43.10
41.38
32.37
33.82
42.17
30.19
35.35
68.56
61.47
56.72
51.82
Fl417
11,861
2
3
4
5
9.90
10.97
10.40
11.03
12,994.83
14,309.41
14,806.25
15,081.88
19,207.45
19,140.47
19,705.09
20,633.11
16,214.79
16,761.76
17,464.90
17,881.99
9.56
20.64
24.83
27.16
61.94
61.37
66.13
73.96
90.44
79.36
75.17
72.84
38.06
38.63
33.87
26.04
52.38
40.73
41.30
46.80
63.29
58.68
52.75
49.24
Pr1002
259,045
2
3
4
5
21.54
21.84
22.07
18.02
306,292.97
317,395.30
345,069.13
360,241.58
369,643.84
375,577.45
390,673.04
404,670.15
336,460.31
352,951.65
367,582.75
380,772.15
18.24
22.53
33.21
39.07
42.69
44.99
50.81
56.22
81.76
77.47
66.79
60.93
57.31
55.01
49.19
43.78
24.46
22.46
17.60
17.15
70.12
63.75
58.10
53.01
ARTICLE IN PRESS

13
Table 7
Experimental results of VV population seeding method.
G Model
ARTICLE IN PRESS
14
Table 8
Performance order for average computation time.
Random Generation
GeneBank
VE
Instance
Performance order (best worst)
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
Random VE VV EV GB NN
Random VE VV EV GB NN
Random VE EV VV NN GB
VE Random VV EV NN GB
Random VE EV VV GB NN
Random EV GB VE VV NN
Random Generation
GeneBank
VE
90
Average Convergence ( %)
S. no.
Nearest Neighbour
EV
VV
70
50
30
10
-10
-30
Nearest Neighbour
EV
VV
-50
25
Eil51
Pr76
KroA100
Pr144
Gil262
Fl417
Pr1002
TSP Instance
20
Time (sec)
Fig. 13. Average convergence of different population seeding techniques.

15
10
0
Eil51
Pr76
KroA100
Pr144
Gil262
Fl417
Pr1002
TSP Instance
Fig. 11. Computation time of different population seeding techniques.
Random Generation
GeneBank
VE
100
Nearest Neighbour
EV
VV
Convergence Diversity ( %)
90
80
70
60
50
40
30
20
10
0
Eil51
Pr76
KroA100
Pr144
Gil262
Fl417
of 5 and 13.71% for Pr1002 with ba value of 2 respectively. On

average, the convergence diversity remains between 20% and 35%
which reveals that the individuals are identical and may stick in
local optimal solution.
The convergence rate and error rate ranges of the population
generated by VE technique are depicted in Figs. 7 and 8 respectively.
4.2.2.2.1. Type 3: VV technique. Table 7 illustrates the experimental results of VV method of ODV based population seeding
technique and the best results are emphasized in bold. These results
may derive the possible observations as follows:
Observation 12: The maximum and minimum values of the best
convergence rate obtained are 92.33% for instance Pr76 with ba
value of 2 and 81.76% for instance Pr1002 with ba value 2 respectively. Thus, at least 82% of convergence has been achieved at the
population initialization stage of the GA for solving TSP problem.
Observation 13: The convergence diversity of the individuals in
the population is maintained between the ranges of 2050% which
is most suitable for exploring search space effectively. The maximum and minimum values obtained are 67.36% for instance of
Pr144 with ba value of 5 and 17.15% for Pr1002 with ba value of
5 respectively.
The convergence rate range and error rate range of the population generated by VV technique are shown in Figs. 9 and 10
respectively.
Pr1002
TSP Instance
Fig. 12. Convergence diversity of different population seeding techniques.
Table 9
Performance order for convergence diversity.
S. no.
Instance
Performance order (high low)
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
Random VV EV GB NN VE
Random EV NN VV VE GB
Random VV VE EV GB NN
Random VE VV EV NN GB
Random VV EV GB VE NN
Random VV EV GB NN VE
Observation 10: The convergence rate of the best individuals

created using the VE population seeding technique remains equal
despite the value of ba. This shows that the individuals generated
suffers a cyclic repetition problem though at least 72.65% of convergence has been achieved at the population initialization stage.
Observation 11: For each instance, the convergence diversity of
individuals increases with ba value. The maximum and minimum
values obtained are 61.83% for instance of Pr144 with ba value
Table 10
Performance order for average convergence.
S. no.
Instance
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
EV VV VE NN GB Random
VV EV VE GB NN Random
VE EV VV GB NN Random
EV VV VE NN GB Random
NN VV EV VE GB Random
VV EV VE NN GB Random
VE VV EV NN GB Random
Table 11
GA conguration parameters.
S. no.
Parameter
Value/technique
1
2
3
4
5
6
7
8
Population size
Generation limit
Crossover method
Crossover probability
Mutation method
Mutation probability
Elitism
Termination condition
100
200
Ordered crossover
0.6 [3,35]
Swap Mutation [35]
0.02 [3,32]
True (3 individuals)
Generation Limit
G Model
ARTICLE IN PRESS

Optimal
Nearest Neighbour
EV
VV
Optimal
Nearest Neighbour
EV
VV
Random Generation
GeneBank
VE
850
Random Generation
GeneBank
VE
200000
700
650
600
550
500
180000
Fittness Value
750
Fittness Value
Fittness Value
800
160000
140000
120000
450
400
100000
3
' ba ' value
Random Generation
GeneBank
VE
Optimal
Nearest Neighbour
EV
VV
Random Generation
GeneBank
VE
Optimal
Nearest Neighbour
EV
VV
4500
22000
90000
80000
4000
3500
70000
3000
60000
2500
50000
Fittness Value
24000
Fittness Value
5000
110000
100000
Random Generation
GeneBank
VE
20000
18000
16000
14000
12000
10000
2000
4
c. KroA100
120000
Random Generation
GeneBank
VE
' ba ' value
b. Pr76
Optimal
Nearest Neighbour
EV
VV
Optimal
Nearest Neighbour
EV
VV
' ba ' value
a. Eil51
Fittness Value
42000
40000
38000
36000
34000
32000
30000
28000
26000
24000
22000
20000
15
' ba ' value
' ba ' value
d. Pr144
' ba ' value
e. Gil262
f. Fl417
Optimal
Nearest Neighbour
EV
VV
Random Generation
GeneBank
VE
Fittness Value
550000
500000
450000
400000
350000
300000
250000
2
' ba ' value
g. Pr1002
Fig. 14. Fitness graph for TSP instances using different population seeding techniques.
Random Populaton
GeneBank
VE
125
Nearest Neighbour
EV
VV
105
4.2.3. Result analyses and discussion

In this section, the effectiveness of Random, NN, GB, EV, VE and
VV population seeding techniques are discussed w.r.t. four performance criteria as follows.
Time (sec)
85
65
45
25
5
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
TSP Instance
Fig. 15. Computation time of different methods analyzed.
Average
4.2.3.1. Computation time. The total time taken for EV, VE and VV
population seeding methods is completely depends on the number of cities in the instance and number of ba value generated
for creating each individual. As a result, VV technique takes more
computation time than EV method and it is further decreased in VE
method. The total number of times ba value generated for EV, VE
and VV methods is (n 1), n ba and n o respectively, where n is
the number of cities of corresponding instance, ba is the best adjacent number and o is the population size. The comparison in terms
of computation time of proposed and existing population seeding
techniques with different TSP instances is shown in Fig. 11. The
G Model
ARTICLE IN PRESS
16
Table 12
Experimental results of random population seeding method.
S. no.
TSPLIB instance
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
Optimal length
426
108,159
21,282
58,537
2378
11,861
259,045
Average tness
499.94
133,483.31
25,443.33
74,833.01
2840.79
14,505.18
335,942.32
Average time
Error rate (%)
Conv. rate (%)
7.74
11.43
15.97
22.13
39.79
71.03
167.41
16.45
20.10
18.60
26.25
18.02
22.14
27.66
83.55
79.90
81.40
73.75
81.98
77.86
72.34
Average time
Error rate (%)
Conv. rate (%)
10.93
16.14
22.56
31.25
56.18
100.29
236.38
9.24
11.50
9.65
9.85
13.64
11.60
11.73
90.76
88.50
90.35
90.15
86.36
88.40
88.27
Table 13
Experimental results of NN population seeding method.
S. no.
TSPLIB instance
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
Optimal length
426
108,159
21,282
58,537
2378
11,861
259,045
Average tness
467.96
121,067.36
23,708.35
64,367.21
2800.92
14,000.61
294,051.26
Table 14
Experimental results of GB population seeding method.
S. no.
TSPLIB Instance
Optimal Length
Average tness
Average Time
Error Rate
Conv. Rate
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
426
108,159
21,282
58,537
2378
11,861
259,045
507.44
126,919.34
24,786.43
67,806.50
2869.72
13,988.40
315,000.77
10.02
14.79
20.67
28.63
51.48
91.89
216.59
14.42%
16.54%
14.80%
14.99%
18.57%
16.64%
16.76%
85.58%
83.46%
85.20%
85.01%
81.43%
83.36%
83.24%
Table 15
Experimental results of EV population seeding method.
S. no.
TSPLIB instance
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
Optimal length
426
108,159
21,282
58,537
2378
11,861
259,045
Average tness
458.19
117,443.94
23,054.95
63,541.92
2678.76
13,654.68
291,997.80
Average time
Error rate (%)
Conv. rate (%)
9.17
13.53
18.91
26.20
47.11
84.10
198.22
5.12
6.08
5.67
6.78
10.38
10.63
11.93
94.88
93.92
94.33
93.22
89.62
89.37
88.07
Average time
Error rate (%)
Conv. rate (%)
7.43
10.96
15.32
21.22
38.16
68.12
160.55
18.78
20.80
17.74
19.50
22.72
20.89
21.00
81.22
79.20
82.26
80.50
77.28
79.11
79.00
Average time
Error rate (%)
Conv. rate (%)
10.86
16.02
22.39
31.02
55.78
99.57
234.68
1.02
3.22
4.13
3.81
5.80
7.11
8.18
98.98
96.78
95.87
96.19
94.20
92.89
91.82
Table 16
Experimental results of VE population seeding method.
S. no.
TSPLIB instance
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
Optimal length
426
108,159
21,282
58,537
2378
11,861
259,045
Average tness
518.83
136,677.89
25,527.27
70,895.65
2979.44
15,316.37
324,520.74
Table 17
Experimental results of VV population seeding method.
S. no.
TSPLIB instance
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
Optimal length
426
108,159
21,282
58,537
2378
11,861
259,045
Average tness
431.73
112,775.72
22,771.79
61,076.87
2555.70
12,906.91
284,340.72
G Model
ARTICLE IN PRESS

Random Populaton
GeneBank
VE
40
Nearest Neighbour
EV
VV
35
Error Rate (%)
30
25
20
15
17
Table 20
Performance order for convergence rate.
S. no.
Instance
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
VV EV NN VE GB Random
VV EV NN GB Random VE
VV EV NN GB VE Random
VV EV NN Random GB VE
VV NN EV GB VE Random
10
5
0
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
Average
TSP Instance
Fig. 16. Error rate of different methods analyzed.
Random Populaton
GeneBank
VE
Nearest Neighbour
EV
VV
Convergence Rate (%)
100
95
90
85
80
Fig. 12, it is understood that the VV technique outperform other

techniques except the random method, which perform better
than VV technique. The performance order of all population seeding techniques in respect of convergence diversity is depicted in
Table 9. The NN and VE methods have the least convergence diversity, which prevent them from evolving optimal solution and stuck
with the local optimal solution. The EV method has better convergence diversity can produce the individuals with high potential
sequence, but the experimental results show that it lacks in exploring the search space completely, which resulted in poor ability
to produce optimal solution sometime. The VV method has better convergence diversity than the other two proposed methods,
which proves that it could explore the problem search space effectively and has a high possibility of nding the optimal solution for
problem of any size.
75
70
65
60
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
Average
TSP Instance
Fig. 17. Convergence rate of different methods analyzed.
Table 18
Performance order for computation time.
S. no.
Instance
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
VE Random EV GB VV NN
Random VE EV GB NN VV
VE Random EV GB NN VV
VE Random EV GB VV NN
Random VE EV GB VV NN
Random VE EV GB VV NN
Random VE EV GB NN VV
performance order of each population seeding technique in respect

of computation time is shown in Table 8.
4.2.3.2. Convergent diversity. The convergence diversity is a property that elucidates the distribution of good and bad quality
individuals among the population generated. Fig. 12 shows the
comparative performance of proposed and existing techniques
w.r.t. convergence diversity for the benchmark TSP instances. From
Table 19
Performance order for error rate.
S. no.
Instance
1
2
3
4
5
6
7
Eil51
Pr76
Kroa100
Pr144
Gil262
Fl417
Pr1002
VV EV NN VE GB Random
VV EV NN Random GB VE
VV NN EV GB VE Random
4.2.3.3. Average convergence. Average convergence explores the

quality of population generated by nding the average of tness of
individuals. Fig. 13 shows the comparison between the proposed
and existing techniques in terms of average convergence. This gure illustrates that the average convergence of the population of EV,
VE and VV methods tends to be equal for large sized problem space
besides with slight variations for small sized problems. The possibility of generating identical individuals is very high in VE method
when compared with other two methods. The performance order
of each population seeding technique in respect of average convergence is shown in Table 10. From the table, it could be observed
that the proposed VV, EV and VE techniques offer better average
convergence rate compared to Random, NN and GB techniques.
4.2.3.4. Best tness value. The best tness value refers to the actual
outcome of the optimal solution. i.e. it refers to the destination
of the solution. The performance comparison in terms of the tness value of the best solution obtained using random, NN, GB,
EV, VE and VV (with different ba values) population seeding techniques w.r.t. the known optimal tness value is shown in Fig. 14.
Fig. 14(a)(g) refers to the performance of the above said techniques
for the benchmark TSP instances Eil51, Pr76, KroA100, Pr144,
Gil262, Fl417 and Pr1002 respectively. The EV method produced
the near optimal solution for Eil51, KroA100 and Pr144 instances
with ba value of 4, 2, and 2 respectively. The near optimal solutions
for Pr76, Gil262, Fl417 and Pr1002 instances are produced by VV
method with ba value 3, 2, 2 and 2 respectively. From the graphs,
it is understood that the VV method outperforms the other techniques. In addition to that it is also observed that the EV method
generated equally good solution several times, by generating best
solution whose tness is near the optimal value, but not for all the
instances. As to conclude, the EV method of population seeding
works well for small sized TSP instances and VV method perform
better for large sized TSP instances compared to other techniques.
4.3. Phase II: Experimentation and result analyses
As said earlier, in this phase, the scope of the experiments is
extended to cover the whole life cycle of the Genetic Algorithm.
G Model
18
ARTICLE IN PRESS
Fig. 18. Convergence graph for TSP instances using different population seeding techniques.
This extended scope of the experiments facilitates to assess the

overall impact of the proposed techniques compared with random,
NN and GB techniques. The crossover and mutation operator used
are Ordered Crossover (OX) [3,5153] and Swap Mutation operator
[30,34,38] respectively. The ordered crossover has been found to
be one of the best in terms of quality and speed to solve TSP and
its working principle is exemplied in [3]. The GA parameters and
the corresponding values are depicted in Table 11. Elitist strategy
is followed to ensure that the ttest individuals in each generation

is carried out to the next generation in order to avoid the replacement of best t individuals with poor individuals in the successive
generations.
For each technique, the executions are carried out for 50 times
with different ba values (ba values are not applicable for random,
NN and GB methods) and the average of 50 runs of each case has
been considered for experimental analyses. The execution proceeds
G Model
ARTICLE IN PRESS
until the termination condition is satised and the tness values of

the converged solutions are used. Tables 1217 shows the Performance of the Random, NN, GB, EV, VE and VV population seeding
methods respectively. The analyses are carried out based on the
performance factors as discussed in Eqs. (17) and (18). The comparative analyses of various population seeding methods in terms
of computation time, error rate and convergence rate are shown in
Figs. 1517 respectively.
Analysis based on computation time elucidates that Random
and VE methods take less time than EV and VV methods, which
are better than GB and NN methods. The order of performance in
terms of computation time for different TSP instances is given in
Table 18. Both convergence and error rates based assessments show
that VV method leads to evolve better convergence with less error
for all the TSP instances despite of problem size. The performance
orders for error and convergence rates are given in Tables 19 and 20
respectively. The VE and NN methods that posses less population
diversity records worst convergence rate for all the TSP instances,
which justify that they have stuck in local optimal solution. The
convergence graph visualizes the improvement in the tness of
the solution by each generation. The convergence graphs for the
TSP instances Eil51, Pr76, Kroa100, Pr144, Gil262, Fl417 and Pr1002
are shown in Fig. 18(a)(g) respectively. From these graphs, the tness improvement of the solutions of VV seeding method converges
better than EV, which outperforms all the other techniques.
19
TSPLIB [54]. The experiments in this research are carried out in

two different phases. The scope of the experiments in the Phase I
is limited to the initialization phase alone, whereas this restricted
scope helps to assess the performance of the proposed technique in
its intended phase alone. In Phase II, the scope of the experiments
is extended to cover the whole life cycle of the Genetic Algorithm,
whereas this extended scope facilitates to assess the overall impact
of the proposed technique. In both the phases, the performance
of the proposed technique has been investigated and compared
with the existing techniques like Random Initialization technique,
Nearest Neighbor Technique and Gene Bank Technique. The experimental analysis based on the performance factors like convergence
and error rate shows that the VV performs better than EV with relatively high computation time compared to existing techniques. This
work can be extended to design an ODV based crossover operator
that exclusively exploit the potential of individuals generated to
improve the performance further, to accomplish 100% convergence
in lesser computation time.
Acknowledgements
This work is a part of the Research Projects sponsored under
the Major Project Scheme, UGC, India, Reference Nos: F. No. 40258/2011(SR), dated 29 June 2011 and F. No. 41-639/2012(SR)/Dt.
16-07-2012. The authors would like to express their thanks for the
nancial supports offered by the Sponsored Agency.
4.4. Discussion
Appendix A. Supplementary data
In summary, the investigations in the Phase I shows that the
VV population seeding method offers better convergence diversity
and initial convergence rate than the EV which is better than NN,
GB, VE and random techniques. It is also noted that, the computation time of VV method is relatively more than the random, EV
and VE methods. The Phase II investigation proves that the VV and
EV methods of population seeding outperforms the random, GB and
NN, which are claimed to be the better population seeding methods.
Thus, the required characteristics of population seeding like randomness, individual diversity and potential sequence are assured
in the proposed ODV based VV and EV population seeding methods. Though, the VV method works best, it is necessary to point
that it could not achieve the complete 100% of convergence for any
of the test instances after certain point of generations. The tness
value of best solution generated by the VV method at the population seeding stage for instance Eil51 is 91.04%, for which the tness
of the best solution after 200 generations of crossover is 98.98%.
This implies that the OX operator does not completely exploit the
potential sequence of individuals generated at the initial stage.
Therefore, genetic operators like edge-recombination and partition
crossover [59,60], which perform well with local optima solutions,
can be used to make use of potential solutions generated using ODV
population seeding technique. And also, designing an ODV based
crossover operator that exclusively exploit the advantages of individuals generated using ODV methods may enable to attain the
100% of convergence, particularly for large sized TSP instances.
5. Conclusion
In this paper, an efcient Ordered Distance Vector (ODV) based
population seeding technique has been proposed for GA to enhance
its overall performance. The proposed technique consists of three
different methods namely EV, VE and VV, which generate a population of individuals with characteristics such as randomness,
diversity and potential sequence. The Traveling Salesman Problem
(TSP) has been chosen as the testbed and the experiments are performed on different sized TSP benchmark datasets obtained from
Supplementary material related to this article can be found, in

the online version, at http://dx.doi.org/10.1016/j.jocs.2013.05.009.
References
[1] M. Albayrak, N. Allahverdi, Development a new mutation operator to solve the
traveling salesman problem by aid of genetic algorithms, Expert Systems with
Applications (July (38)) (2011) 13131320, ISSN:0957-4174.
[2] K. Helsgaun, An effective implementation of the LinKernighan traveling
salesman heuristic, European Journal of Operational Research 126 (1) (2000)
106130.
[3] S.R. Shubhra, B. Sanghamitra, K.P. Sankar, Genetic operators for combinatorial optimization in TSP and microarray gene ordering, Journal
of Applied Intelligence 26 (June (3)) (2007) 183195, Kluwer Academic
Publishers.
[4] J.K. Lenstra, A.H.G. Rinnooy Kan, Some simple applications of the traveling salesman problem, Operational Research Quarterly 26 (1975)
717733.
[5] M. Grtschel, M. Jnger, G. Reinelt, Optimal control of plotting and drilling
machines: a case study, Mathematical Methods of Operations Research 35
(January (1)) (1991) 6184.
[6] R.D. Plante, T.J. Lowe, R. Chandrasekaran, The product matrix traveling salesman problem: an application and solution heuristics, Operations Research 35
(1987) 772783.
[7] R.E. Bland, D.E. Shallcross, Large traveling salesman problem arising from
experiments in X-ray crystallography: a preliminary report on computation,
Operations Research Letters 8 (3) (1989) 125128.
[8] H.D. Ratliff, A.S. Rosenthal, Order-picking in a rectangular warehouse: a solvable case for the travelling salesman problem, Operations Research 31 (1983)
507521.
[9] L. Fiechter, A parallel tabu search algorithm for large travelling salesman problems, Discrete Applied Mathematics and Combinatorial Operations Research
and Computer Science 51 (3) (1994) 243267.
[10] M. Zachariasen, M. Dam, Tabu search on the geometric traveling salesman
problem, in: Proc. International Conference on Metaheuristics, 1995, pp.
571587.
[11] B. Freisleben, P. Merz, Genetic local search algorithm for solving symmetric and
asymmetric traveling salesman problems, in: Proceedings of the IEEE Conference on Evolutionary Computation, IEEE Press, Nagoya, 1996, pp. 616621.
[12] J.Y. Potvin, Genetic algorithm for the traveling salesman problem, Annals of
Operations Research 63 (1996) 339370.
[13] C.F. Tsai, C.W. Tsai, Yang T, A modied multiple-searching method to genetic
algorithms for solving traveling salesman problem, IEEE International Conference on Systems, Man, and Cybernetics 3 (2002) 69.
[14] S. Elaoud, J. Teghem, T. Loukil, Multiple crossover genetic algorithm for
the multiobjective traveling salesman problem, in: Electronic Notes in
G Model
20
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
ARTICLE IN PRESS
Discrete Mathematics, Elsevier, Issue 36, May, ISSN:1571-1653, 2010, pp.

939946.
S.S. Ray, S. Bandyopadhyay, S.K. Pal, New operators of genetic algorithms
for traveling salesman problem, ICPR, Cambridge, UK, 2004, ICPR-04, 2:
497500.
M. Albayrak, N. Allahverdi, Development a new mutation operator to
solve the Traveling Salesman Problem by aid of Genetic Algorithms, in:
Expert Systems with Applications, Issue 38, July, ISSN:0957-4174, 2011 July,
pp. 13131320.
P.C. Chang, W.S. Huang, C.J. Ting, Dynamic diversity control in genetic algorithm
for mining unsearched solution space in TSP problems, Expert Systems with
Applications 37 (3) (2010) 18631878.
M. Dorigo, L.M. Gambardella, Ant colony system: a cooperative learning
approach to the traveling salesman problem, IEEE Transactions on Evolutionary
Computation 1 (1) (1997) 5356.
T. Stutzle, M. Dorigo, ACO algorithms for the traveling salesman problem, evolutionary algorithms in engineering and computer science, John Wiley and Sons,
1999, ISBN 978-0-471-99902-7.
I. Ellabib, P. Calamai, O. Basir, Exchange strategies for multiple ant colony system, Information Sciences 177 (5) (2007) 12481264.
L. Li, S. Ju, Y. Zhang, Improved ant colony optimization for the traveling
salesman problem, in: Proceedings of the 2008 International Conference
on Intelligent Computation Technology and Automation. Vol. 1, 2008, pp.
7680.
X. Shi, L. Wang, Y. Zhou, Y. Liang, An ant colony optimization method for prizecollecting traveling salesman problem with time windows, in: Proceedings of
the Fourth International Conference on Natural Computation, Jinan, China. Vol.
7, 2008, pp. 480484.
P. Victer Paul, T. Vengattaraman, P. Dhavachelvan, Improving efciency
of peer network applications by formulating distributed spanning tree,
in: Proceedings 3rd International Conference on Emerging Trends in
Engineering and Technology, ICETET 2010. Art. no. 5698439, 2010, pp.
813818.
J.Y. Potvin, The traveling salesman problem: a neural network perspective,
ORSA Journal on Computing 5 (4) (1993) 328348.
M. Saadatmand-Tarzjan, M. Khademi, T.M. R. Akbarzadeh, H.A. Moghaddan, A
novel constructive-optimizer neural network for the traveling salesman problem, IEEE Transactions on Systems, Man, Cybernetics Part B: Cybernetics 37
(4) (2007) 754770.
X. Geng, Z. Chen, W. Yang, D. Shi, K. Zhao, Solving the traveling salesman problem based on an adaptive simulated annealing algorithm with greedy search,
Applied Soft Computing 11 (June (4)) (2011) 36803689.
X.F. Xie, J. Liu, Multiagent optimization system for solving the traveling salesman problem (TSP), IEEE Transactions on Systems, Man, and Cybernetics Part
B: Cybernetics 39 (2) (2008) 489502.
Y. Marinakis, M. Marinaki, A hybrid genetic particle swarm optimization
algorithm for solving the vehicle routing problem, Expert Systems with Applications 37 (2) (2010) 14461455.
S.M. Chen, C.Y. Chien, A new method for solving the traveling salesman problem based on the genetic simulated annealing ant colony system with particle
swarm optimization techniques, in: Proceedings of the 2010 International Conference on Machine Learning and Cybernetics, Qingdao, Shandong, China, 2010,
pp. 24772482.
Shyi-Ming Chen, Chih-Yao Chien, Solving the traveling salesman problem based
on the genetic simulated annealing ant colony system with particle swarm
optimization techniques, in: Expert Systems with Applications, Elsevier, Issue
38, May, ISBN:0957-4174, 2011, pp. 1443914450.
G. Andal Jayalakshmi, S. Sathiamoorthy, A hybrid genetic algorithm: a new
approach to solve traveling salesman problem, International Journal of Computational Engineering Science 2 (2) (2001) 339355.
W. Yingzi, H. Yulan, G. Kanfeng, Parallel search strategies for TSPs using a greedy
genetic algorithm, in: Third International Conference on Natural Computation,
ICNC 2007, Vol. 3, August, 2007, pp. 786790.
W. Pullan, Adapting the Genetic Algorithm to the Travelling Salesman Problem,
in: The 2003 Congress on Evolutionary Computation, CEC03, Vol. 2, 2003, pp.
10291035.
K. Katayama, H. Sakamoto, H. Narihisa, The efciency of hybrid mutation
Genetic Algorithm for the Travelling Salesman Problem, in: Mathematical and Computer Modelling, Elsevier, Issue 31, October, 2000, pp.
197203.
P. Dhavachelvan, G.V. Uma, Complexity measures for software systems:
towards multi-agent based software testing proceedings 2005, in: International Conference on Intelligent Sensing and Information Processing, ICISIP05
2005, Art. no. 1529476, 2005, pp. 359364.
O. Yugay, I. Kim, B. Kim, F.I.S. Ko, Hybrid genetic algorithm for solving Traveling
Salesman Problem with sorted population, in: Third IEEE International Conference on Convergence and Hybrid Information Technology (ICCIT 2008), 2008,
ISSN:978-0-7695-3407.
P. Dhavachelvan, G.V. Uma, Multi-agent framework construction for intra class
testing of object-oriented software, in: 18th ISCIS 2003, Springer Verlag Lecture Notes in Computer Science (LNCS), ISSN:0302-9743, Vol. 2869, 2003, pp.
992999.
L. Fuyan, C. Chouyong, L. Shaoyi, An improved genetic approach,
in: International Conference on Neural Networks and Brain,
2005, ICNN&B05, Vol. 2, Beijing, ISBN:0-7803-9422-4, 2005, pp.
641644.
[39] W.A. Chang, R.S. Ramakrishna, A genetic algorithm for shortest path routing
problem and the sizing of populations, IEEE Transactions on Evolutionary Computation 6 (December (6)) (2002) 566579, ISSN:1089-778X.
[40] V. Togan, A.T. Daloglu, An improved genetic algorithm with initial
population strategy and self-adaptive member grouping, Journal of
Computers and Structures 86 (June (1112)) (2008) 12041218, ISSN:
0045-7949.
[41] S. Venkatesan, P. Dhavachelvan, C. Chellapan, Performance analysis of mobile
agent failure recovery in e-service applications, in: International Journal of
Computer Standards and Interfaces, Vol. 32, Issue 12, Elsevier, ISSN:09205489, 2005, pp. 3843.
[42] C.F. Tsai, C.W. Tsai, Yang T, A modied multiple-searching method to genetic
algorithms for solving traveling salesman problem, IEEE International Conference on Systems, Man, and Cybernetics 3 (2002) 69.
[43] G. Reinelt, The traveling salesman: computational solutions for TSP
applications, Lecture Notes in Computer Science (1994) 840, Springer
Verlag.
[44] J.L. Bentley, Fast algorithms for geometric traveling salesman problems, ORSA
J Computing 4 (4) (1992) 387411.
[45] D.S. Johnson, L.A. McGeoch, The Traveling Salesman Problem: A Case Study
in Local Optimization. Local Search in Combinatorial Optimization, Wiley and
Sons, New York, 1996.
[46] Y. Lu, S. Lu, F. Fotouhi, Y. Deng, S. Brown, Incremental genetic K-means algorithm and its application, Gene Expression Data Analysis. Submitted to BMC
Bioinformatics (2004).
[47] X. Hue, Genetic Algorithms for Optimization: Background and Applications,
Edinburgh Parallel Computing Centre, Univ. Edinburgh, Edinburgh, Scotland,
Ver 1.0, 1997 February.
[48] J. Lu, B. Feng, B. Li, Finding the optimal gene order for Genetic Algorithm, in:
Fifth World Congress on Intelligent Control and Automation, Vol. 3, WCICA
2004, June, ISBN:0-7803-8273-0, 2004, pp. 20732076.
[49] Z. Jianxin, C.T. Beijing, Solving TSP with novel local search heuristic genetic algorithms, in: Fourth International Conference on Natural Computation, ICNC08,
Vol. 1, Jinan, ISBN:978-0-7695-3304-9, 2008, pp. 670674.
[50] L.-Y. Wang, Z. Jie, H. Li, An improved genetic algorithm For TSP, in: International
Conference on Machine Learning and Cybernetics, Vol. 2, Hong Kong, August,
ISBN:978-1-4244-0973-980, 2007, pp. 925928.
[51] P. Larranaga, C. Kuijpers, R. Murga, I. Inza, Dizdarevic S, Genetic algorithms for
the traveling salesman problem: a review of representations and operators,
Articial Intelligence Review 13 (1999) 129170.
[52] Goldberg DE, Genetic Algorithm in Search, Optimization and Machine
Learning, Machine Learning, Addison-Wesley, New York, 1989, pp.
188.
[53] S.S. Ray, S. Bandyopadhyay, S.K. Pal, New Operators of Genetic Algorithms for
Traveling Salesman Problem, ICPR, Cambridge, UK, 2004, pp. 497500, ICPR-04
2.
[54] http://comopt.i.uni-heidelberg.de/software/TSPLIB95/ (accessed 21.09.12).
[55] J.H. Holland, Adaptation in Natural and Articial Systems, The University of
Michigan Press, Ann Arbor, MI, 1975.
[56] S.S. Ray, S. Bandyopadhyay, S.K. Pal, Genetic operators for combinatorial optimization in TSP and microarray gene ordering, Applied Intelligence 26 (3)
(2007) 183195.
[57] H. Maaranen, K. Miettinen, M.M. Makela, Quasi-random Initial Population for
Genetic Algorithms, Computers and Mathematics with Applications 47 (12)
(2004) 18851895.
[58] S. Lin, B.W. Kernighan, An effective Heuristic Algorithm for the Traveling Salesman Problem, Operations Research 21 (2) (1973) 498516.
[59] D. Whitley, D. Hains, A. Howe, Tunneling between optima: partition
crossover for the traveling salesman problem, in: Proceedings of the 11th
ACM Annual Conference on Genetic and Evolutionary Computation, 2009,
pp. 915922.
[60] D. Whitley, D. Hains, A.E. Howe, A hybrid genetic algorithm for the traveling salesman problem using generalized partition crossover, PPSN 1 (2010)
566575.
[61] J.S. Lawrence, M.M. Amini, Performance characteristics of alternative genetic
algorithmic approaches to the traveling salesman problem using path representation: an empirical study, European Journal of Operational Research (April
(108)) (1998) 551570, Elsevier, ISSN: 0377-2217/98.
[62] L. Qu, R. Sun, A synergetic approach to genetic algorithms for solving traveling
salesman problem, Information Sciences (July (117)) (1999) 267283, Elsevier,
ISSN:0020-0255.
[63] P. Dhavachelvan, G.V. Uma, Reliability enhancement in software testing: an
agent-based approach for complex systems, in: 7th ICIT 2004, Springer Verlag
Lecture Notes in Computer Science (LNCS), Vol. 3356, ISSN:0302-9743, 2004,
pp. 282291.
[64] L.-N. Xing, Y.-W. Chen, K.-W. Yang, F. Hou, X.-S. Shen, H.-P. Cai, A
hybrid approach combining an improved genetic algorithm and optimization
strategies for the asymmetric traveling salesman problem, in: Engineering
Applications of Articial Intelligence, Elsevier, Issue 21, April, ISSN:0952-1976,
2008, pp. 13701380.
[65] P. Victer Paul, N. Saravanan, S.K.V. Jayakumar, P. Dhavachelvan, R. Baskaran, QoS
enhancements for global replication management in peer to peer networks,
Future Generation Computer Systems 28 (3) (2012) 573582.
[66] P. Dhavachelvan, G.V. Uma, Multi-agent based framework for intra-class testing
of object-oriented software, International Journal on Applied Soft Computing
5 (2) (2005) 205222, Elsevier.
G Model
ARTICLE IN PRESS
P. Victer Paul is a research scholar, pursuing Ph.D. in the

Department of Computer Science, Pondicherry University, Pondicherry, India. He has completed his B.Tech. in
information technology from SMVEC and M.Tech. in network and internet engineering, Pondicherry University,
Pondicherry, India. Currently he is working in the elds
of evolutionary computing and distributed systems.
A. Ramalingam is working as the associate professor in

the Department of Computer Applications, Sri Manakula
Vinayagar Engineering College, Pondicherry, India. He has
completed his MCA and M.Phil. in computer science. He
also completed M.Tech. in information technology in 2004
and currently he is pursuing his Ph.D. in Pondicherry Engineering College, Pondicherry, India. He is having around
20 years of experience in teaching. Currently he is working
in the elds of genetic algorithms and software engineering.
Dr. R. Baskaran is currently associate professor, Department of Computer Science and Engineering, Anna
University. He completed his B.Tech. in electrical and
electronics engineering (2000) and M.Tech. in computer
science and engineering (2001). He obtained his Ph.D. in
computer science and engineering in 2007 from Anna University, Chennai. He has around a decade of experience
in the education and research. His research areas include
evolutionary computing, service computing and multimedia databases. He has published more than 60 research
articles in International & National Journals, Conferences
and Books. He is the member of various National and
International bodies like The Institution of Electronics and
Telecommunication Engineers, Computer Society of India (CSI), International Network for Engineering Education and Research, International Association of Engineers
and International Congress for Global Science and Technology.
21
Dr. P. Dhavachelvan is working as the professor, Department of Computer Science, Pondicherry University, India.
He completed his B.Tech. in electrical and electronics engineering in Madras University (1997), India. He obtained
his M.Tech. in computer science and engineering (2000)
and Ph.D. in computer science and engineering (2007)
from Anna University, Chennai. He is having around 15
years of experience as an Academician, Researcher and
Administrator. Presently he is heading the Department of
Computer Science, Pondicherry Central University, India.
His research areas include Software Engineering, Web
Service Computing and Evolutionary Algorithms. As the
Main and Coauthor, he has more than 100 publications in
his credit. The publication list includes National and International Journals, National
and International Conferences, Books and Book Chapters.
Dr. K.Vivekanandan is working as the senior professor
in the Department of Computer Science and Engineering, Pondicherry Engineering College, India. He obtained
his B.E. degree in electronics and communication engineering from Coimbatore Institute of Technology in 1986
and M.Tech. in computer science and engineering from
Indian Institute of Technology, Bombay (India) in 1991. He
received his Ph.D. in computer science and engineering
from Pondicherry University, India in 2005. He is having around 21 years of experience in teaching, research
and administration. He worked in Indian Institute of Science, Bangalore and Central Research Laboratory, Bharat
Electronics Limited, Bangalore. Since 1992 he has been
working in Pondicherry Engineering College. His research interest includes Software Engineering, Web Service Computing and Information Security. In his credit,
he is having around 50 National and International publications in various forums
like Journals, Conferences and Books.
Dr. R. Subramanian is the senior professor in the
Department of Computer Science, Pondicherry Central
University, India. He completed his B.Sc. in mathematics in the Madurai Kamaraj University, India in 1982. He
received his M.Sc. and Ph.D. in mathematics from Indian
Institute of Technology, Delhi, India in 1984 and 1989
respectively. He is having around 23 years of experience
in teaching and research. As a part of administration, he
had been the HOD of the Department of Computer Science
and currently he is the Dean of the School of Engineering
and Technology, Pondicherry Central University, India. His
specialization includes parallel and distributed systems,
robotics and evolutionary algorithms. He has published
more than 50 National & International Journal & Conference publications, Books
and Book Chapters.

J3 - Elsevier JOCS - A New Population Seeding Technique For Permutation-Coded Genetic Algorithm

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

J3 - Elsevier JOCS - A New Population Seeding Technique For Permutation-Coded Genetic Algorithm

Uploaded by

Copyright:

Available Formats

G Model

JOCS-208; No. of Pages 21

Contents lists available at SciVerse ScienceDirect

Journal of Computational Science

A new population seeding technique for permutation-coded Genetic

Department of Computer Science, Pondicherry University, Puducherry, India

Corresponding author. Tel.: +91 0413 2274430.

congurations encourage researchers to design novel GA with a

JOCS-208; No. of Pages 21

P. Victer Paul et al. / Journal of Computational Science xxx (2013) xxxxxx

Find Best individual in Population (Bestindiv)

New Population (Gen=Gen+1) with

Fig. 1. The generic ow of Genetic Algorithm.

and mechanisms of natural selection. In 1975, John Holland

Evaluate the tness value of each individual based on a tness

The search space for the TSP is a set of permutation of n cities.

Check whether number of generation exceeds the generation

d(c i , c (i+1) )

d(c i , c (i+1) )

Such that, 1 (n + 1) and  {} 

JOCS-208; No. of Pages 21

P. Victer Paul et al. / Journal of Computational Science xxx (2013) xxxxxx

3.1. Problem statement

where, d(ci , cj ) refers to the distance between the cities ci and cj .

Ordered Distance Vector based population seeding technique

2 (ci ) 2 (c(i+1) ) 2 (c(i+2) ) . . . 2 (cn )

where k is the kth permutation of n cities, k = (1, 2, 3, . . ., o). o

3.3. ODV based population seeding techniques

Rank of distance of city cx with cy denoted as R(cx , cy ) refers to

The rank of distance of the city is used to choose the city at

3.3.1. Type 1: Equi-begin with Variable diversity (EV)

Equal ba diversity (Ed): An integer value bax is generated

Based on these two characteristic factors, there are three

2 (1) 2 (c2 ) 2 (c3 ) . . . 2 (cn )

3.3.2. Type 2: Vari-begin with Equal diversity (VE)

Rank of distance of cities

Suitable crossover operators

Suffer cyclic repetition

JOCS-208; No. of Pages 21

P. Victer Paul et al. / Journal of Computational Science xxx (2013) xxxxxx

The PODM generated using this type can be represented as:

solution with minimum convergence time. The maximum number

where The rst city remains same for ba number of individuals

variables used in the proposed algorithm is shown in Fig. 3. The

initial population. This procedure gets n and DM as inputs and

JOCS-208; No. of Pages 21

P. Victer Paul et al. / Journal of Computational Science xxx (2013) xxxxxx

Best tness Worst

procedure, each row from the DM is retrieved and sorted, the

generated using VE method and returns the generated population

JOCS-208; No. of Pages 21

P. Victer Paul et al. / Journal of Computational Science xxx (2013) xxxxxx

JOCS-208; No. of Pages 21

P. Victer Paul et al. / Journal of Computational Science xxx (2013) xxxxxx

Fig. 5. The range of variation in convergence rate of EV method.

Fig. 6. The range of variation in error rate of EV method.

(ii) Average convergence: It is the average of convergence rate of

Average tness Optimal tness

where Average tness is the average tness value of solutions in

Fitness Optimal tness

where CRBest Indiv refers to the convergence rate of best solution

4.1.3. Experimental phases

that have been chosen for experimentation are Eil51, Pr76,

d(c i , c (i+1) )

d(c i , c (i+1) )

Such that, 1 (n + 1) and {}