A COMPARISON BETWEEN SWARM INTELLIGENCE ALGORITHMS FOR ROUTING PROBLEMS

Electrical & Computer Engineering: An International Journal (ECIJ) Volume 5, Number 1, March 2016
DOI : 10.14810/ecij.2016.5102 17
A COMPARISON BETWEEN SWARM INTELLIGENCE
ALGORITHMS FOR ROUTING PROBLEMS
Shima Sabet1
, Mohammad Shokouhifar*2
, and Fardad Farokhi1
1
Department of Electrical Engineering, Islamic Azad University of Central Tehran
Branch, Tehran, Iran
2
Department of Electrical Engineering, ShahidBeheshti University G.C., Tehran, Iran
ABSTRACT
Travelling salesman problem (TSP) is a most popular combinatorial routing problem, belongs to the class
of NP-hard problems. Many approacheshave been proposed for TSP.Among them, swarm intelligence (SI)
algorithms can effectively achieve optimal tours with the minimum lengths and attempt to avoid trapping in
local minima points. The transcendence of each SI is depended on the nature of the problem. In our studies,
there has been yet no any article, which had compared the performance of SI algorithms for TSP perfectly.
In this paper,four common SI algorithms are used to solve TSP, in order to compare the performance of SI
algorithms for the TSP problem. These algorithms include genetic algorithm, particle swarm optimization,
ant colony optimization, and artificial bee colony. For each SI, the various parameters and operators were
tested, and the best values were selected for it. Experiments oversome benchmarks fromTSPLIBshow that
artificial bee colony algorithm is the best one among the fourSI-basedmethods to solverouting problems
like TSP.
KEYWORDS
Travelling salesman problem (TSP), Swarm intelligence algorithms, Genetic algorithm (GA), Ant colony
optimization (ACO), Particle swarm optimization (PSO), Artificial bee colony (ABC).
1. INTRODUCTION
Travelling salesman problem (TSP) is the most popular routing problem. It denotes a travelling
salesman who wants to visit a number of nodes (cities) exactly once, and finally returns to the
starting node (city). Objective of the problem is to determine a tour with the minimum
distance.TSP is a sub-problemof many applications such as transportation [1],
wirelesscommunications [2], vehicle routing [3], integrated circuits[4] and semiconductor
industries [5].
In recent years, many approaches have been developed forTSP, whichcan be classified into
exactand approximate solutions [6]. Exact approaches, e.g., Branch & Bound [7], Branch &Cut
[8], and Dynamic Programming [9], can only be performed for very small datasets. Therefore,
researchers have to utilize approximate methods for the real-world problems. These approaches
can find a set of near-optimal solutions in a reasonable time, rather than an exact algorithm that
guarantees to achieve the optimal solution but in an exponential time. The operation ofan exact
method is to generate all possible tours, andselect the bestone with the minimum path length.The
number ofall possible permutations of ܰ cities isܰ!, where each tour can be represented in 2ܰ
different manner according to the initial selected city and the direction of the tour. So the size of

18
the search space can be calculatedaccording to Eq. (1). It is obviously that evaluation of all the
ሺܰ − 1ሻ!/2solutions is practically impossible in term of computational time, even for20 cities
[10].
ܶ =
ே!
ଶ∗ே
=
ሺேିଵሻ!
ଶ
(1)
Approximate approaches are included of heuristic and meta-heuristic algorithms such as
simulated annealing [11], tabu search [12],and swarm intelligence (SI) algorithms [13-25]. SI
algorithms are population-based meta-heuristic algorithms, which have random-based global
search strategies, and have attracted a great deal of attentions [26-32]. They able to carry out
better search process toward other algorithms, and do not require the gradient information.These
methods find a set of near-optimal solutions in a reasonable time rather than exact algorithms
which can find the optimum solution but in exponential time. In recent years, the SI algorithms
effectively have been developed for TSP, such as: genetic algorithm (GA) [13-17], particle swarm
optimization (PSO) [18-20], ant colony optimization (ACO) [21-23] and artificial bee colony
(ABC) [24-25].
In [13], a GA-based algorithm has been introduced for TSP. It uses a topological encoding
schemeto form ordering of vertices in a directed graph. Also, Albayrak, et. al[17] have proposed a
greedy sub tour mutation (GSTM) operator to improve the convergence of GA. It can efficiently
search locally within the previous solutions like the other mutation operators.
The traditional PSO was proposed for the continuous problems. However some modifications of
PSO have been applied for TSP. Wang, et. al [18] modified the PSO using the concepts of
Adjustment Operator and Adjustment Sequence. In this algorithm, the particles not only adjust
it’s own flying speed and the best individual of the swarm, but also learn from other individuals
according to proposed certain probability. Additionally, the mutation of velocity was used. Shi, et.
al[20] modified the subtraction operator between two particle positions in PSO. The crossover
eliminated technique and two local search techniques were also executed their method. These
local search techniques were used to accelerate the convergence speed, by employing the
generalized chromosome.
Shang, et. al [21] combined ACO and association rule (AR) as a hybrid swarm intelligence based
algorithm for the TSP. It is worth noting that the AR has been presentedto findthe best data
sequence in data mining. Li, et. al [22] also presented an improved ACO, in which, a modified
probability-based selection scheme is utilized by using Held-Karp lower bound in order to have a
better trade-off between the influence of pheromone trail and heuristic information.
In [24] a novel ABC algorithm has been suggested to solve TSP so that it mixtures the route
construction mechanismwith path relining (PR). This method introduces attributes of the guiding
solution into the initial solution to obtain high quality solutions as quickly as possible. The other
ABC-based algorithm is a Hybrid Mutation-based ABC algorithm (named HMABC) to solve TSP
[25], which employed the hybrid mutation scheme to search the neighbourhood of the found
solutions, in order to achieve better solutions.
As mentioned above, many SI-based methods have been proposed for TSP. In this paper, we aim
to carry out a comparison between SI algorithms, in order to evaluate the performance of SI
algorithms for the routing problems like TSP. The rest of the paper is organized as follows: In

19
section 2, the proposed framework of each EA for TSP has introduced. In section 3, experimental
results are shown. And finally a comparison between SI algorithms for TSP is described in section
4.
2. SWARM INTELLIGENCE BASED SOLUTIONS FOR TSP
A routing solution in the TSP can be represented as a graph ‫ܩ‬ = ሺܸ, ‫ܧ‬ሻ, in which,ܸ =
{1, 2, . . . , ܰ} is the set of all nodes (cities) within the problem graph, and ‫ܧ‬ = {ሺ݅, ݆ሻ|݅, ݆ܸ߳} is the
set of all possible edges between thenodes. More specifically, each node represents the position of
a city, whereas each edge corresponds to a connection path between two cities. The distance ݀௜௝
which is associated with edge ሺ݅, ݆ሻ, represents the Euclidean distance from city ݅ to city ݆, and is
calculated according to Eq. 2. The heuristic information is calculated before employing the SI
algorithms at offline step. So, the distances of all edges were saved.
݀௜௝ = ටሺ‫ݔ‬௜ − ‫ݔ‬௝ሻଶ + ሺ‫ݕ‬௜ − ‫ݕ‬௝ሻଶ(2)
In all SI algorithms, iteratively after construction the solutions, the qualities of the gathered
solutions are evaluated by the defined objective function, which is same for all SI algorithms. It
can beevaluated for the each solution (a chromosome, a particle, an ant, or a bee) by simply
calculating the sum ofEuclidean distances of the consecutive edges within the tour as follows:
‫ݐݏ݋ܥ‬௞ = ∑ ݀௟
ே
௟ୀଵ (3)
To solve TSP by SI algorithms, at the first, an initial population is generated to search among N-
dimension search space according to get an optimum tour.In order to generate the initial solutions,
at first, a random city isselected as the initial node(e.g.,city ݅).Then, the remained cities are
consecutively added to the tour, until all cities will be selected. At the each step, the next city is
chosen according to a defined probability which corresponds to the reverse of its distances from
the current city. We called it the probability-based nearest neighbour (PNN) rule.
2.1. Proposed GA-based Algorithm for TSP
Genetic algorithm (GA) was invented firstly by Holland in 1975 [33]. In order to solve TSP via
GA, each chromosome can be encoded bypermutation coding to represent a tour. On the other
hand, each gene withina chromosome is a positive integer which means the ID of a city thatthe
salesman is passing from that.
At first, the initial population of GA is randomly constructedby thePNN rule, in such a way
thatnorepetitive city occurs within each chromosome.At the every iteration,quality of each
chromosome is evaluated according to the predefined objective function in Eq. 3. As a result, all
solutions are sorted from the best to the worst, according to the objective function.Then, some of
the best ones are selected as parents based on a selection strategy, e.g., roulette wheel selectionor
elitism selection. After performing the crossover operation, the mutation operator is applied for
some gathered offspring, to change randomly in the value of some genes.

Electrical & Computer Engineering: An International J
A. Crossover
In order togenerate an offspring via
the best individuals. There is an essential problem in the crossover step for solving the TSP: with
a traditional crossover operator, when two correct parents are recombined, each of the resultant
offspring is malformed. Because after the crossover, some repetitive
offspring. Some methods were introduced aim to overcome this problem: Sheng
greedy crossover operation (GX). Majumdar
(MCX), modified order crossover (MOX),
had applied a partially matched crossover (PMX)
copy-first principle. In the permutation encoding, the PMX and MOX crossover operators have
been efficiently used.
In the MOX operator, some gene
positions.The rest of genes are taken in the same
utilize the MOX operator with three crossover pointsat random positions. For the first offspring,
the genes are copied from the first parent till the initial crossover point, also from the second point
till the third point. Then the genes of the second parent are scanned, and the
second parent which are not within the first offspring, will be appear in the first offspring with the
same order that they are came in the second parent. Illustration of MOX operator can be seen in
Fig. 1. Also, the PMX operator works as fo
positions. The second parts of theparents are exchanged,
does not appear in the second part,
remained numbers are appeared
second parent. The operation of the
Figure 1.Illustration of the
Figure 2.Illustration of the
& Computer Engineering: An International Journal (ECIJ) Volume 5, Number 1, March 2016
In order togenerate an offspring via crossover operator, two parents are randomly chosen among
There is an essential problem in the crossover step for solving the TSP: with
offspring is malformed. Because after the crossover, some repetitive cities are appears in the
offspring. Some methods were introduced aim to overcome this problem: Sheng et al.
greedy crossover operation (GX). Majumdaret al. [14] proposed modified cycle crossover
modified order crossover (MOX), and sequence crossover(SX) operators. Choi
had applied a partially matched crossover (PMX).Also, Xing et al. [34] proposed the same
first principle. In the permutation encoding, the PMX and MOX crossover operators have
genes of the first parent are printed in the offspring, in the same
are taken in the same sequence order as in the second parent.
with three crossover pointsat random positions. For the first offspring,
till the third point. Then the genes of the second parent are scanned, and the numbers of the
Fig. 1. Also, the PMX operator works as follows: two crossover points are chosen at random
he second parts of theparents are exchanged, and then, each number of first parent that
appear in the second part, is written in the first offspring in the same position
ed in the first offspring with the same sequence order as
operation of the PMX mechanism can be seen in Fig. 2.
the MOX crossover operatorwith threepoints on random positions.
the PMX crossover operatorwith twopoints on random positions.
, March 2016
20
, two parents are randomly chosen among
There is an essential problem in the crossover step for solving the TSP: with
cities are appears in the
et al. [15] used a
ed modified cycle crossover
. Choi et al. [16]
] proposed the same-site-
first principle. In the permutation encoding, the PMX and MOX crossover operators have
, in the same
order as in the second parent. We
with three crossover pointsat random positions. For the first offspring,
numbers of the
llows: two crossover points are chosen at random
each number of first parent that
same position.The
order as in the
operatorwith threepoints on random positions.
operatorwith twopoints on random positions.

B. Mutation
In TSP, there are various mutation operators like
[13-17]. Also, we have introduc
neighbour change (NC) in [25], as
genesare randomly selected and exchanged
random location within the offspring.T
same sequence. The NS and the NSC operations can be
Figure 3.Illustration of the NS mutation operation
Figure 4.Illustration of the NSC mutation
Also, the NC operatorworks as follow:
• One city is randomly selectedas the initial city of the tour
• A city must be selected in the neighborhood of the city
distance, has more chance to be selected. The probability ofcity
of its distance from the current city
• The positionof the city ݅
• The simple shift-changeis performed for thetwo selected cities.
C. Overall Steps in Proposed GA
In this paper, the MOX crossover operator and the NC mutation operator were employed. The
proposed GA-based algorithm for TSP can be described as follow:
Step 1: Initialization
• Set the parameters.
• Set the max-iterations.
Step 2: Generation of initial population by PNN rule.
• Each chromosome selects a random tour based on PNN rule and constructs a solution.
Step 3: Evaluation of the gathered solutions by Eq.3
Step 4: Check the stop criterion:
• Stop, if the maximum number of iterations
In TSP, there are various mutation operators like swap, inversion, routeand shift-c
introduced neighbourswap (NS), neighbourshift-change (NSC), and
], as some mutation operators for TSP. In the NS, two
and exchanged together. The NSCselects two neighbour
location within the offspring.Then, it shifts all remained genes between them with the
. The NS and the NSC operations can be illustrated inFigures 3 and 4, respectively
Figure 3.Illustration of the NS mutation operation.
Figure 4.Illustration of the NSC mutation operation.
the NC operatorworks as follow:
randomly selectedas the initial city of the tour (e.g., the city ݅).
A city must be selected in the neighborhood of the city ݅, in such a way thateach
distance, has more chance to be selected. The probability ofcity ݆is proportional to the diverse
urrent city.
and the lately selected city are determined in the solution.
geis performed for thetwo selected cities.
Overall Steps in Proposed GA
based algorithm for TSP can be described as follow:
Step 2: Generation of initial population by PNN rule.
ach chromosome selects a random tour based on PNN rule and constructs a solution.
Step 3: Evaluation of the gathered solutions by Eq.3
umber of iterations reached.Otherwise, continue.
, March 2016
21
change (RSC)
hange (NSC), and
mutation operators for TSP. In the NS, two neighbour
neighbour genes at
shifts all remained genes between them with the
, respectively.
in such a way thateach has less
is proportional to the diverse
and the lately selected city are determined in the solution.
ach chromosome selects a random tour based on PNN rule and constructs a solution.

Step 5: Generation of new population. These 3 actions are done until completion the population:
• Selection: two best chromosomes are chosen as parents by Elitism strateg
• Crossover: MOX operator is applied to generate two offspring from theparents.
• Mutation: the NC operator is done with a defined probability for some offspring.
Step 6: Go to step 3 and continue.
2.2. Proposed ACO-based Algorithm for TSP
Ant colony optimization (ACO) proposed by Dorigo [3
inspired by the swarm basedbehaviour
chemical material (called pheromone
idea should be used to mimic the
problem graph, as shown in Fig. 5.
Figure 5.Graph representation in
example it chooses the city a3,a6 , a
ant comes back to
A. Construction the suitable solutions
Given a colony of k artificial ants; these ants perform a number of iterations
each ant starts from an initial
according to the probability rule, until
the city ito selectthe city jas its next city iscalculat
( ) , ,k
ij kP t
where,‫ܩ‬௞is the set of all conditional cities
߬௜௝and ߟ௜௝ are the pheromone
respectively. Here,ߟ௜௝ is calculated as
and j. On the other hand, α is a parameter in the range
importance between the pheromone trail and
: two best chromosomes are chosen as parents by Elitism strategy.
: MOX operator is applied to generate two offspring from theparents.
: the NC operator is done with a defined probability for some offspring.
based Algorithm for TSP
timization (ACO) proposed by Dorigo [35] and firstly applied to solve TSP. ACO
behaviour of real ants to find the shortest paths.A
called pheromone), when they walking. In order to perform ACO for TSP,
to mimic thebehaviourof real ants withartificial ones walking around the
as shown in Fig. 5.
TSP: ant is firstly at node a1and should add anext city to its path. For
a4 , a5 , a2 , respectively, based on transition probability rule. F
ant comes back to the initial citya1 and completes its tour.
Construction the suitable solutions
artificial ants; these ants perform a number of iterations.At the every
city, randomly.Then, it consecutively selects the next
according to the probability rule, until it completes the tour. The probability ofi-th ant
as its next city iscalculated as follows [36]:
( ) .
( ) , ,
( ) .
ij ij
l jij k
il il
l
G
t
P t
t
α
α
τ η
τ η
∈=
∑
(4)
conditional cities which have notbeen selected yet by the
are the pheromone amount and the heuristic information for choosing city
is calculated as 1/݀௜௝, where ݀௜௝ is the Euclidean distance between cities
α is a parameter in the range of [0,1] which define
the pheromone trail and the heuristic information.
, March 2016
22
: MOX operator is applied to generate two offspring from theparents.
: the NC operator is done with a defined probability for some offspring.
] and firstly applied to solve TSP. ACO
nts deposit a
ACO for TSP, an
walking around the
next city to its path. For
respectively, based on transition probability rule. Finally, the
very iteration,
the next cities
th ant located at
the i-th ant, and
choosing city ݆,
is the Euclidean distance between cities i
fine the relative

23
B. Pheromone updating
At the every iteration,all antsshould be evaluated by means of the objective function.Then, two
high-quality ants are chosen for the pheromone deposition on their edges: the first is the best ant
in the current iteration, and the second is the global best ant found so far. The pheromone
updating is done as follows:
1 for edges of the best ant
otherwise
D
(1 ) .
0
ij ij
t t
τ λ τ+ 
= − + 

(5)
where߬௜௝
௧
and ߬௜௝
௧ାଵ
are the volume of pheromone on edge (݅, ݆) respectively at iteration ‫ݐ‬ and at the
next iteration. Also, D is aconstant parameter determining thepheromone deposition value, and λ
is the amount of the evaporation factor.
C. Overall Steps in Proposed ACO
The main steps of proposed ACO-based algorithm can be summarized as follows:
Step 1: Initialization and set the parameters.
Step 2: Generation of initial solutions by PNN rule.
• Each ant selects a random tour based on PNN rule, and constructs a solution.
Step 3: Evaluation of achieved solution by ants, using Eq. 3.
• Stop, if the maximum number of iterations reached.Otherwise, continue.
Step 5: Pheromone updating.
• Evaporation: It is done for the all edges of the graph.
• Deposition: It is performedonly for the edges of the two best ants.
Step 6: Generation of new solutions according to Eq. 4.
Step 7: Go to step 3.
2.3. Proposed PSO-based Algorithm for TSP
Particle swarm optimization (PSO) introduced by Eberhart and Kennedy in 1995 [37], inspired by
social behaviour of bird flocking and fish schooling. In the TSP, the position of each particle can
beillustratedas ܺ௜ = (‫ݔ‬௜ଵ, ‫ݔ‬௜ଶ, … , ‫ݔ‬௜஽), which defines apossible travelling path represented by the
݅-th particle with the city sequence of‫ݔ‬௜ଵ → ‫ݔ‬௜ଶ → ⋯ → ‫ݔ‬௜஽. The velocity of each particle is
represented as ܸ௜ = (‫ݒ‬௜ଵ, ‫ݒ‬௜ଶ, … , ‫ݒ‬௜஽). Here, ݅ = 1,2, … , ݉, where݉ is thenumber of all particles,
and ‫ܦ‬isthe number of all cities within the problem graph. The velocities of the particles which are
far from an own optimum or the general optimum, can sharply rise, but the velocities of high-
quality particles change slightly.
The PSO is initialized randomly by generation a population of solutions called particles. The
qualities of the gathered particles are evaluated by the predefined objective function in Eq. 3.The
movementof each particle at the every iterationis influenced by theown best experience (called
‘pbest’), andthe global best solutionfound byall particlesso far (called ‘gbest’). On the other hand,
each particle is updated by influence of the both pbest and gbest. Positions and velocities of
particles dynamically are updated according to following equations:

24
ܸ௜ௗ(‫ݐ‬ + 1) = ‫ݓ‬௜. ܸ௜ௗ(‫)ݐ‬ + ‫ܥ‬ଵ.‫.݀݊ܽݎ‬ ቀ‫݌‬௕௘௦௧೔೏
(‫)ݐ‬ − ‫ݔ‬௜ௗ(‫)ݐ‬ቁ + ‫ܥ‬ଶ. ‫.݀݊ܽݎ‬ (݃௕௘௦௧೒೏
− ‫ݔ‬௜ௗ(‫))ݐ‬ (6)
‫ݔ‬௜ௗ(‫ݐ‬ + 1) = ‫ݔ‬௜ௗ(‫)ݐ‬ + ‫ݒ‬௜ௗ(‫ݐ‬ + 1) (7)
whereC1 and C2 are two constant parameters so called acceleration coefficients. Also,randisa
random number uniformly distributed in the range of[0,1]. Moreover, wi is the inertia weight,
often decreasing linearly from Wmax to Wmin during execute algorithm that tunes as Eq. 8,
whereItmaxis the maximum number of iteration, and It is the current iteration.
‫ݓ‬௜ = ‫ݓ‬௠௔௫ + (‫ݐܫ/ݐܫ‬௠௔௫). (‫ݓ‬௠௔௫ − ‫ݓ‬௠௜௡) (8)
In order to solve TSP via PSO, each particle should be represented as a stringof positive integers.
So the classical updating scheme for velocities and positions (Eqs. 6 and 7) cannot be used for
the TSP, and a mechanism must be employed in order to iteratively update the positions toward
the pbest and gbest.
In this work, we propose a hybrid crossover-based scheme to combine the current position of the
particle with its pbest and the gbest. In this way at the first, the two strings which represent the
position of pbestandgbest are combined together using the PMX crossover operation, and
generate the Velocity. Then the current particle moves toward the Velocity using the MOX
crossover operator. Finally, the new generated string is considered as the particle positionfor the
next iteration. The overall steps of proposed PSO-based algorithm can be summarized as
follows:
Step 1: Initialization and set the parameters.
Step 2: Generation of initial particles by PNN rule.
Step 3: Evaluation of the particles by Eq. 3.
Step 4: Determination of pbest of each particle, and gbest.
Step 5: Generation of the new population:
• The pbest of each particle is combined with the gbest using the PMX crossover operator,
and the Velocity of that particle is made.
• Each particle moves toward the bothpbest andgbest, by combination of the particle with
it’sVelocity using the MOX crossover operator.
Step 7: Go to step 3.
2.4. Proposed ABC-based Algorithm for TSP
Artificial bee colony (ABC) is proposed by Karaboga in 2006 [38] which inspired of the
intelligent foraging behaviour of honey bees. In ABC, the colony consists of three groups of bees:
employed bees, onlookers and scouts [39]. Each employedgoes onto the previously visited food
source area in her memory and achieves a newfood source. Provided that the nectar (fitness) of
the new solution be higher than the previous one, the employed bee forgets the previous solution
and memorizes the new one.Otherwise, she keeps the previous. Each onlooker gives the

25
information of employed bees, and makes decision to selectan employed through the shared
information from the dances of the employed bees within the hive. Also,those employed bees
which were abandoned, become scout, and start randomly search among the search space, as same
as the initial solution generation.
In order to solve TSP via ABC,a food source represents a feasible tour.Also, the nectar of the food
source corresponds to the quality of the tour. Initially, a set of food sources are randomly
generated by employed bees. Then, the nectar (fitness) of the gathered food sources are
determined by means of the objective function. As a result, the qualityof the initial solutions
isevaluated.
D. Selection of the best employed bees
After all employed bees have completed their search and constructed their solutions, they come
back into the hive and share theinformation about the quality of thefood sources with the
onlookers waiting within the hive. The probability of݇-th employed bee to be selected by an
onlooker beecan be expressed as follows:
1
k
k N
i i
N
P
N
α
α
=
=
∑
(9)
whereܰ௞ is the amount of nectar within food source k, and ∑ ܰ௞
ே
௜ୀଵ is the total amount of
achieved nectar. ܰ௞is considered as 1/‫ݐݏ݋ܥ‬௞. Increasing in the value of α (α>1) leads to select the
better bees as scouts.
E. Recruitment process for onlooker bees
After selecting the best employed bees, the recruitment process shouldbe performed, in such a
way that some onlookers are assigned for the each best employed bee. The more quality the
employed exists, the more onlookers would be assigned. The number of selected onlookers for the
each of the best employed can be calculated asܵ௜ = ߚ × (ܲ௜/‫)ܪ‬ + ߛ, where ܵ௜ is the percentage of
the onlookers selected for the ݅-thbest employed bee. ܲ௜is the preference number of the ݅-th
employed, and H is the number of all selected best employed bees. Also, ß and γare two constants
which determine the recruitment scheme.If ß=0, the number of selected onlookers for the allbest
employed beesare determined uniformly. The larger ß, the more onlookers are selected for the
higher quality employed bee.
F. Neighbourhood search mechanism
In this step, each onlooker andits assigned employed beego onto the neighbourhood area of the
food source which has beenfound by the employed bee, in the previous iteration. In order to
generate a new solution in the vicinity of the current one, some parts of the current tour should be
changed. In order to do that, we proposed a hybrid mutation scheme in [25] for searching the
neighbourhood of each scout bee. In this way, three mutation operators consist of NSC, NS and
NC are used, which each of them has the own advantage and prevents the search process from a
kind local minima. Each bee uses the NC operator with the %50 probability, and the chance of
applying the NS and the NSC operators are %20 and %30 respectively.

26
G. Overall Steps in Proposed ABC
As mentioned above, the main steps of the ABC-based routingcan be summarized as follows:
Step 1: Initialization and set the parameters of ABC.
Step 2: Generation of initial solutions by PNN rule.
• Each employed bee construct a solution according to PNN rule.
Step 3: Evaluation of the achieved solutions by Eq. 3.
Step 4: Updating of theglobal best solution found so far.
Step 5: Selection of the bestemployed bees by onlookers, and performing the recruitment process.
Step 7: Neighbourhood search using thehybrid mutation-based local search scheme.
3. SIMULATION RESULTS
3.1. Simulation Settings
All of the experiments tested on a PC with 2.53GHZ
processor and 4GB
memory running MATLAB
R2012b in windows 8. The experiments were carried out on benchmark problems from TSPLIB
[40] as seen in Table 1. The first column is the name ofproblems, the second is the number of
cities within the problem, and the third is the best tour length.To adjust theoperators
andcontrollable parameters of the each SI algorithm, different operators and values were
evaluated and the best ones were considered for the final simulations. The population size and the
maximum number of iterations were set to the same value of 100 and 500, respectively, for the all
SI algorithms.
Table 1.Benchmarks.
Problem Number of Cities Optimal Solution
At48 48 10628
Eil51 51 426
Berlin52 52 7542
Eil76 76 538
KroA100 100 21282
Lin105 105 14379
Beir127 127 118282
KroA200 200 29368
Tuning the GA:.In order to perform GA, the PMX and MOX, NSC,NS andNCoperators are
evaluated as the crossover and mutation operators. With respect to the obtained results, the MOX
crossover operator and the NC mutation operatorwere selected for applying GA in our
experiments. Setting the parameters of GA can be summarized in Table 2.
Table 2.Parameter setting for the proposed GA-based algorithm.
Parameter Value
Max Iteration 500
Number of Chromosomes 100
Selection Type Elitism Selection

27
Parameter Value
Number of Selected Parents 5
Crossover Operator MOX
Number of Points for Crossover 3 points
Max distance of two neighbours for applying mutation Number of Cities/5
Mutation operator NC
Mutation probability 0.5/ Number of Cities
Tuning the ACO:The pheromone of all edges is considered within the range of [0,1], where the
initial pheromonewas set to 0.5.Different values were evaluated for the evaporation factor λand
finally it was set to λ=0.15. The pheromone depositionon the edges of the two selected best ants
was defined asD=0.5. The parameter setting for ACO can be shown in Table 3.
Table 3.Parameter setting for the proposed ACO-based algorithm.
Parameter Value
Max Iteration 500
Number of Ants 100
Number of Candida ants 2
Initial Pheromone of All Edges 0.5
Evaporation factor (λ) (Eq. 5) %15
Deposition amount (D) (Eq. 5) 0.5
α (Eq. 4) 2
Max Iteration 500
Tuning the PSO: For PSO, the three point MOX crossover was used on random positions, in
order to combine the pbest and gbest. And finally the three point MOX crossover was performed
to combine the particle with the pbest and gbest, aim to move the particle toward them.
Table 4.Parameter setting for the proposed PSO-based algorithm.
Parameter Value
Max Iteration 500
Number of Particles 100
Type of Combination of pbest and gbest to Generate
the Velocity
PMX Operator
Type of Combination the Particle with itsVelocity MOX Operator
Tuning the ABC:.As mentioned above, a hybrid mutation scheme was used, in order to search
the neighbourhood of the solutions found by the scouts. We tested some combination of mutation
operators in our previous work [31]. Respect to the obtained results, we employed the NS, NSC,
and NC operators with their corresponding probability.
Table 5.Parameter setting for the proposed ABC-based algorithm.
Parameter Value
Max Iteration 500
Number of Employed Bees 100
Number of Scout Bees 5
Max distance of two neighbours for applying mutation Number of Cities/5
Probability of applying NS operator %20
Probability of applying NSC operator %30
Probability of applying NC operator %50

28
Parameter Value
β -20
γ 32
H 5
α 2
3.2. Simulation Results
We performed each EA 10 times. The best achieved solution, and the average for 10 runs, using
each algorithm can be shown in Table 6. The first column is the problem names; the second, third,
fourth and fifth columns are consist of the best found tour length, and the average for 10 runs,
respectively using GA, ACO, PSO, and ABC. Moreover, comparison of the running time of the
different algorithms can be seen in Table 7. The best solution achieved by ABC for some
problems is shown in Figures 6 to 9. In these Figures, the x-axis and y-axis describe the location
of cities. Each point is corresponding with each city.
Table 6. Comparison of the average and best error rate for 10 runs (in percentage).
Problem GA ACO PSO ABC
Best Avg. Best Avg. Best Avg. Best Avg.
At48 1.23 2.16 0.72 1.64 0.67 0.92 0.31 0.82
Eil51 1.71 3.52 0.43 2.88 0.2 1.37 0 0.77
Berlin52 1.62 2.41 0 2.22 1.03 1.91 0.27 0.91
Eil76 2.03 4.64 0 2.93 0 1.15 0 1.15
KroA100 1.92 3.29 1.25 3.78 1.37 2.56 1.08 1.67
Lin105 2.67 4.17 1.08 3.92 0.93 2.57 0.36 1.72
Beir127 3.04 4.45 2.12 4.89 1.25 3.41 1.25 2.12
KroA200 4.33 6.45 2.45 5.78 1.87 3.61 1.37 2.08
Table 7.Comparison of the running time (in seconds).
Problem GA ACO PSO ABC
At48 35 53 45 31
Eil51 38 61 49 35
Berlin52 38 65 53 32
Eil76 42 102 84 29
KroA100 52 155 119 36
Lin105 57 168 123 44
Beir127 68 213 175 45
KroA200 107 366 288 94

Figure 6.The Best Solution for Berlin52, Found by ABC.
Figure 7.The Best Solution for At48, Found by ABC.
Figure 6.The Best Solution for Berlin52, Found by ABC.
Figure 7.The Best Solution for At48, Found by ABC.
, March 2016
29

Figure 8.The Best Solution for Lin105, Found by ABC.
Figure 9.The Best Solution for KroA100, Found by ABC.
Figure 8.The Best Solution for Lin105, Found by ABC.
9.The Best Solution for KroA100, Found by ABC.
, March 2016
30

31
4. CONCLUSION
In this paper, we compared the performance of the four famous SI algorithms include of GA,
PSO, ACO, and ABC to solve TSP. Respect to the obtained results, ABC achieved better
performance compared with the other SI algorithms, especially for high dimension problems.
ACO has memory, in which, the ants that were optimal in the past, are tugged to return towards
due to their deposited pheromone. However, as the number of cities be increased, the performance
of ACO is reduced, because of complexity in running time. So, it is recommended for the
problems with fewer than 80 cities. Although GA and PSO have very good performance for the
class of parameter optimization problems, but they have some limitations for the permutation
problems such TSP: the traditional scheme of GA and PSO lead to generate some repetitive cities
in their solutions. Although some modifications have been utilized to handle this problem,these
algorithms cannot compete withACO and ABC. The ABC algorithm has a strong global search
capability in the problem space and can efficiently find optimal tour. For some of the high
dimensional problems, the other SI algorithms often trapped in local minima points after finding a
sub-optimal solution. However, ABC can efficiently explore in the search space until an optimal
solution found. The ABC algorithm considers both running time and solution quality. ABC is
quicker in finding optimal solution than the other SI algorithms; also it is quite simple and easy to
apply. The performances of other SI algorithms are reduced especially for high dimension
problems, and ABC can be employed to guarantee achieving the optimal solution.
REFERENCES
[1] X. Wang & A. C. Regan, (2002) “Local truckload pickup and delivery with hard time window
constraints”, Transportation Research, Part B, pp. 97–112.
[2] M. Shokouhifar & A. Jalali, (2015) “A new evolutionary based application specific routing protocol
for clustered wireless sensor networks”, AEU-International Journal of Electronics and
Communications, Vol. 69, No. 1, pp. 432-441.
[3] M. Shokouhifar, A. Jalali & H. Torfehnejad, (2015) “Optimal routing in traveling salesman problem
using artificial bee colony and simulated annealing”, In 1st National Road ITS Congress.
[4] M. Shokouhifar & A. Jalali, (2015) “An evolutionary-based methodology for symbolic simplification
of analog circuits using genetic algorithm and simulated annealing”, Expert Systems with
Applications, Vol. 42, No. 3, pp. 1189-1201.
[5] C. A. Silvaa, M. C. Sousaa & T. A. Runkler, (2008) “Rescheduling and optimization of logistic
processes using GA and ACO”, Engineering Applications of Artificial Intelligence, Vol. 21, pp. 343-
352.
[6] M. Shokouhifar & S. Sabet, (2012) “PMACO: A pheromone-mutation based ant colony optimization
for traveling salesman problem”, In 2012 International Symposium on Innovations in Intelligent
Systems and Applications (INISTA), pp. 1-5.
[7] M. Padberg & G. Rinaldi, (1987) “Optimization of a 532-city symmetric traveling salesman problem
by branch and cut”, Operational Researches, Vol. 6, pp.1–7.
[8] M. W. Padberg & S. Hong, (1980) “On the symmetric traveling salesman problem: a computational
study”, Math. Prog. Stud, Vol. 12, pp. 78–107.
[9] B. E. Bellman, (1963) “Dynamic programming treatment of the traveling salesman problem”, J.
Assoc. Comput. Mach., Vol. 9.
[10] A. Ugur & D. Aydin, (2009) “An interactive simulation and analysis software for solving TSP using
ant colony optimization algorithms”, Advances in Engineering Software, Vol. 40, pp. 341–349.
[11] C. Song, K. Lee & W. D. Lee, (2003) “Extended simulated annealing for augmented TSP and multi-
salesmen TSP”, In Proceedings of the international joint conference on neural networks, pp. 2340–
43.

32
[12] J. L. Ryan, T. G. Bailey, J. T. Moore & W. B. Carlton, (1998) “Reactive Tabu search in unmanned
aerial reconnaissance simulations,” In Proceedings of the 1998 winter simulation conference, pp. 873–
879.
[13] C. Moon, J. Kim, G. Choi & Y. Seo, (2002) “An efficient genetic algorithm for the traveling salesman
problem with precedence constraints,” European Journal of Operational Research, Vol. 140, pp. 606-
617.
[14] J. Majumdara& A. K. Bhunia, (2011) “Genetic algorithm for asymmetric traveling salesman problem
with imprecise travel times,” Journal of Computational and Applied Mathematics, Vol. 235, pp.
3063–3078.
[15] H. Ismkhan & K. Zamanifar, (2012) “Developing Improved Greedy Crossover to Solve Symmetric
Traveling Salesman Problem”, In Proceedings of IEEE conferences.
[16] I. C. Choi, S. I. Kim & H. S. Kim, (2003) “A genetic algorithm with a mixedregion search for the
asymmetric traveling salesman problem,” Computers & Operations Research, Vol. 30, pp. 773–786,
2003.
[17] M. Albayrak & N. Allahverdi, (2011) “Development a new mutation operator to solve the Traveling
Salesman Problem by aid of Genetic Algorithms,” Expert Systems with Applications, Vol. 38, pp.
1313–1320.
[18] C. Wang, J. Zhang, J. Yang, C. Hu & J. Liu, (2005) “A Modified Particle Swarm Optimization
Algorithm and its Application For Solving Traveling Salesman Problem,” In IEEE Conference
Proceedings, pp. 689-694.
[19] S. A. Subbotin & A. A. Oleynik, (2006) “PSO with Control of Velocity Change for Feature
Selection,” In Conference Proceedings, pp. 80-82.
[20] X. H. Shi, Y. C. Liang, H. P. Lee, C. Lu & Q. X. Wang, (2007) “Particle swarm optimization-based
algorithms for TSP and generalized TSP,” Information Processing Letters, Vol. 103, pp. 169–176.
[21] G. Shang, Z. Lei, Z. Fengting & Z. Chunxian, (2007) “Solving Traveling Salesman Problem by Ant
Colony Optimization Algorithm with Association Rule,” Third International Conference on Natural
Computation (ICNC), pp. 1-4.
[22] L. Li, S. Ju & Y. Zhang, (2008) “Improved Ant Colony Optimization for the Traveling Salesman
Problem,”International Conference on Intelligent Computation Technology and Automation, pp. 76-
80.
[23] A. Ugur & D. Aydin, (2009) “An interactive simulation and analysis software for solving TSP using
Ant Colony Optimization algorithms,” Advances in Engineering Software, Vol. 40, pp. 341–349.
[24] X. Zhang, Q. Bai & X. Yun, (2011) “A New Hybrid Artificial Bee Colony Algorithm for the
Traveling Salesman Problem,” In IEEE Conference Proceedings, pp. 155-159.
[25] S. Sabet, M. Shokouhifar & F. Farokhi, (2013) “A hybrid mutation-based artificial bee colony for
traveling salesman problem”, Lecture Notes on Information Theory, Vol. 1, No. 3, pp. 99-103.
[26] M. Shokouhifar & A. Jalali, (2014) “Real-time task scheduling in heterogeneous multiprocessor
systems using artificial bee colony”, In 22th
Iranian conference on electrical engineering (ICEE
2014), pp. 1007-1012.
[27] M. Shokouhifar & F. Farokhi, (2010) “An artificial bee colony optimization for feature subset
selection using supervised fuzzy c_means algorithm”, In 3rd
International conference on information
security and artificial intelligent (ISAI), pp. 427-432.
[28] M. Shokouhifar & A. Jalali, (2015) “Automatic simplified symbolic analysis of analog circuits using
modified nodal analysis and genetic algorithm”, Journal of Circuits, Systems and Computers, Vol. 24,
No. 4, pp. 1-20.
[29] M. Shokouhifar & A. Jalali, (2014) “Automatic symbolic simplification of analog circuits in
MATLAB using ant colony optimization”, 22th
Iranian conference on electrical engineering (ICEE
2014), pp. 407-412.
[30] S. Sabet, M. Shokouhifar & F. Farokhi, (2013) “A discrete artificial bee colony for multiple knapsack
problem”, Int. J. Reasoning-based Intelligent Systems, Vol. 5, No. 2, pp.88–95.
[31] M. Shokouhifar & G. S. Abkenar, (2011) “An artificial bee colony optimization for mri fuzzy
segmentation of brain tissue”, In 2011 International Conference on Management and Artificial
Intelligence IPEDR,Vol. 6.

[32] M. Shokouhifar & A. Jalali, (2016) “Evolutionary based simplified symbolic PSRR analysis of analog
integrated circuits”, Analog Integrated Circuits and Signal Processing
0680-2.
[33] J. Holland, (1975) “Adaptation in natural and artificial
Arbor.
[34] L. N. Xing, Y. W. Chen, K. W. Yang, F. Hou, X. S. Shen & H. P. Cai, (2008) “A hybrid approach
combining an improved genetic algorithm and optimization strategies for the asymmetric traveling
salesman problem,” Engineering Applications of Artificial Intelligence
[35] M. Dorigo, V. Maniezzo & A. Colorni, (1996) “The ant system: Optimization by a colony of
cooperating agents,” IEEE Trans. Syst. Man Cybernet, Part B
[36] M. Shokouhifar & F. Farokhi, (2010) “
with ant colony optimization”,
446.
[37] J. Kennedy & R. Eberhart, (1995) “Particle swarm optimization,”
pp. 1942-1948.
[38] D. Karaboga & B. Basturk, (2007) “A powerful and efficient algorit
optimization: artificial bee colony (ABC) algorithm”,
459-471.
[39] M. Shokouhifar & S. Sabet, (2010) “
networks and artificial bee colony optimization
(ICMV), pp. 502-506.
[40] http://www.iwr.uniheidelberg.de/groups/comopt/software/TSPLIB95/TSPLIB.html
AUTHOR
Mohammad Shokouhifar received his B.S. from Islamic Azad University, Dezfoul,
Iran, in 2008. He received two M.S. degrees from Islamic Azad University of Central
Tehran Branch and ShahidBeheshti University, Tehran, Iran, in 2011 and 2013,
respectively. He is currently a Ph.D. candidate in Electronic Engineering at
ShahidBeheshti University, Tehran, Iran. His research interests include symbolic
analysis and design of analog OTAs, wireless sensor networks, fuzzy sets and
systems and swarm intelligence algorithms.
Shokouhifar & A. Jalali, (2016) “Evolutionary based simplified symbolic PSRR analysis of analog
Analog Integrated Circuits and Signal Processing, DOI: 10.1007/s10470
J. Holland, (1975) “Adaptation in natural and artificial systems” University of Michigan press, Ann
L. N. Xing, Y. W. Chen, K. W. Yang, F. Hou, X. S. Shen & H. P. Cai, (2008) “A hybrid approach
Engineering Applications of Artificial Intelligence, Vol. 21, pp. 1370
M. Dorigo, V. Maniezzo & A. Colorni, (1996) “The ant system: Optimization by a colony of
IEEE Trans. Syst. Man Cybernet, Part B, Vol. 26 pp. 29–41.
hokouhifar & F. Farokhi, (2010) “Feature selection using supervised fuzzy C-means algorithm
”, In 3rd International Conference on Machine Vision (ICMV)
J. Kennedy & R. Eberhart, (1995) “Particle swarm optimization,” In IEEE conference proceedings
D. Karaboga & B. Basturk, (2007) “A powerful and efficient algorithm for numerical function
optimization: artificial bee colony (ABC) algorithm”, Journal of global optimization
M. Shokouhifar & S. Sabet, (2010) “A hybrid approach for effective feature selection using neural
networks and artificial bee colony optimization”, In 3rd international conference on machine vision
http://www.iwr.uniheidelberg.de/groups/comopt/software/TSPLIB95/TSPLIB.html.
Mohammad Shokouhifar received his B.S. from Islamic Azad University, Dezfoul,
in 2008. He received two M.S. degrees from Islamic Azad University of Central
Tehran Branch and ShahidBeheshti University, Tehran, Iran, in 2011 and 2013,
respectively. He is currently a Ph.D. candidate in Electronic Engineering at
ty, Tehran, Iran. His research interests include symbolic
analysis and design of analog OTAs, wireless sensor networks, fuzzy sets and
systems and swarm intelligence algorithms.
, March 2016
33
Shokouhifar & A. Jalali, (2016) “Evolutionary based simplified symbolic PSRR analysis of analog
, DOI: 10.1007/s10470-015-
systems” University of Michigan press, Ann
L. N. Xing, Y. W. Chen, K. W. Yang, F. Hou, X. S. Shen & H. P. Cai, (2008) “A hybrid approach
, Vol. 21, pp. 1370–1380.
M. Dorigo, V. Maniezzo & A. Colorni, (1996) “The ant system: Optimization by a colony of
means algorithm
In 3rd International Conference on Machine Vision (ICMV), pp. 441-
In IEEE conference proceedings,
hm for numerical function
Journal of global optimization, Vol. 39, pp.
A hybrid approach for effective feature selection using neural
In 3rd international conference on machine vision

A COMPARISON BETWEEN SWARM INTELLIGENCE ALGORITHMS FOR ROUTING PROBLEMS

Recommended

More Related Content

What's hot (17)

Similar to A COMPARISON BETWEEN SWARM INTELLIGENCE ALGORITHMS FOR ROUTING PROBLEMS (20)

More from ecij (20)

Recently uploaded (20)

A COMPARISON BETWEEN SWARM INTELLIGENCE ALGORITHMS FOR ROUTING PROBLEMS