SciELO - Scientific Electronic Library Online

vol.22 issue2Improving the Boilerpipe Algorithm for Boilerplate Removal in News Articles Using HTML Tree StructureBrain-Computer Interface (BCI) Development for Motor Disabled People Integration in a Water Fountains Company author indexsubject indexsearch form
Home Pagealphabetic serial listing  

Services on Demand




Related links

  • Have no similar articlesSimilars in SciELO


Computación y Sistemas

On-line version ISSN 2007-9737Print version ISSN 1405-5546

Comp. y Sist. vol.22 n.2 México Apr./Jun. 2018  Epub Jan 21, 2021 


Genetic Algorithm for Solving Multiple Traveling Salesmen Problem using a New Crossover and Population Generation

Dharm Raj Singh1 

Manoj Kumar Singh1 

Tarkeshwar Singh2 

Rajkishore Prasad3 

1 Banaras Hindu University, DST-Centre for Interdisciplinary Mathematical Sciences (DST-CIMS), India

2 Department of Mathematics, Birla Institute of Technology and Science Pilani, K K Birla Goa Campus, India

3 B.N. College, Patna, India


In this paper, we proposed a new crossover operator and a population initialization method for solving multiple traveling salesmen (MTSP) problem in genetic algorithm (GA) framework. The group theory based technique of intital population generation ensures the uniqueness of members in population, hence no redundancy in the search space and also remove the random initialization effect. The new crossover is based on the intuitive idea that the city in sub optimal / optimal tours occurs at same position. In this crossover the Hamming distance is preserved and there is very less chance to generate child same as members in the population, so diversity of the population is not much affected. For efficient representation of search space, we exploited the multi-chromosome representation technique to encode the search space of MTSP. We evaluate and compare the proposed technique with the methods using crossover TCX, ORX +A, CYX +A and PMX +A reported in [35] for two standard objective functions of the MTSP, namely, minimising total travel cost of m tours of the m salesman and minimising the longest tour cast travel by any one salesman. Experimental results show that the GA with proposed population initialization and crossover gives better result compared to all four methods for second objective, however, in very few cases slightly degraded result for first objective.

Keywords: Genetic algorithm; multiple traveling salesman problem; group theory; crossover operator; 2-opt mutation

1 Introduction

The MTSP is an extension of classical traveling salesmen problem (TSP) [1]. In classical TSP, a salesman visits every city exactly only once in fully connected networks of city and return back to start city such that tour cast is minimum. In MTSP, m (>1), salesman travels the fully connected network of cities such that at a time a group of city is visited exactly once by only one salesman out of m salesman and salesman return to start city i.e. tours of each salesman have distinct city. Each salesman must have starting and ending position is same. Given a collection of n cities and the distance (cost), of travel between every pair of city must be partitioned into m tours for m>1 salesmen to serve a set of n>m cities. The objectives of MTSP are:

  1. Sum of the cost of m tours of the m salesman is minimum.

  2. Maximum tour cost travel by any one salesman is minimized.

We have also considered two additional constraints on the solution, similar to Crater et al. [6]: start and end city is the same for all salesmen and this city is called as home city, every salesmen visit at least one city other than home city. The second objective is related to the balancing of workload among the salesman. Clearly, MTSP is NP-Hard as classical TSP (m = 1), which is well known NP-Hard.

A number of real life problem are modeled using MTSP, which is useful for many applications business, industry and engineering. The most common application of MTSP are in the area of scheduling and routing e.g. print press scheduling [14], interview scheduling [12], crew scheduling[2], hot rolling scheduling[34], workload balancing [25], design of global navigation satellite system surveying networks [30], school bus routing [2], and vehicle scheduling problem [23, 27].

In recent years, many heuristic or meta-heuristic algorithms have been developed for solving the NP-hard optimization problems, such as Simulated Annealing [7, 18], Genetic Algorithms [1, 22], Ant Colony Optimization [36] and Particle Swarm Optimization [24]. These methods are also exploited to solve MTSP [4, 5, 6, 21, 32, 33, 35, 36].

Finding global solution for NP-Hard problems within affordable time and computing resources is intractable. Instead of global optimal solution, a suboptimal solution with reasonable computational load is obtained and it will further pruned for better solution using some heuristic about the solution. During recent years, researcher has shown great interest in finding effective algorithms in heuristic framework to solve MTSP and TSP. A comprehensive survey on heuristic methods is available in [19], and references there in. In this paper we focused on the development of algorithm for MTSP in Genetic Algorithm (GA), framework. GA is an iterative, population based, heuristic search technique. Several GA based methods are proposed depending on the different way of representing chromosome and genetic operations.

Tang et al. [34] proposed a GA with one chromosome representation for MTSP to solve the hot rolling production scheduling problem. Malmborg et al. [23] and Park et al [27] used a two chromosome representation in their genetic algorithm based approaches for MTSP and applied it for vehicle scheduling. Carter et al. [6] developed a genetic algorithm for MTSP with his proposed two-part chromosome representation and corresponding genetic operators. Brown et al. [5] proposed a grouping GA that uses a chromosome presentation. Singh et al. [31] proposed a grouping GA, in this they represented chromosome as set of m tours and their proposed crossover and mutation operator. Chromosome representation of Singh et al. [31] is similar to multi chromosome representation, except that tours are not assigned to specific salesmen. Yuan et al. [37] proposed genetic algorithm for MTSP with two-part chromosome and a new crossover, he named it as two-part crossover (TCX).

Novelty of all the aforementioned method is use of new chromosome representation and modified form of genetic operator (crossover, mutation), which was used in classical TSP [37]. But, these chromosome representations have many redundant solutions, i.e., many chromosome seems to be different but they represent the same MTSP solution. The set of all possible chromosomes (representation space / search space), is much larger than the set of all possible solution to the problem (problem space). However, GA is searching based method and it works on representation space, therefore it has to explore the larger space, and hence, the performance of the GA degraded severely. The detail pros and cons of chromosome representations used in these methods are discussed in Section 2.

In this paper, we proposed the genetic algorithm for the MTSP. Our proposed genetic algorithm uses group theory to generate initial population and a proposed new crossover operator. We use muti-chromosome representation scheme for encoding the representation space [1]. In multi-chromosome representation, the size of the search space is same as the size of two-part chromosome representation. However, multi-chromosome encoding is much similar to the characteristics of MTSP, salesmen are separated from each other “physically”, i.e., each salesman visit different city except start and end city. This representation more easily interpretable to the MTSP solution space. Upon use of group theory method for chromosome initialization, the multi chromosomes representation has no redundancy (duplicate solutions) in search space.

We have compared the results our proposed method with the result reported in Yuan et al. [38] using two part chromosome crossover (TCX), ordered crossover operator with an asexual crossover (ORX +A), cycle crossover operator with an asexual crossover (CYX +A), and the partially-matched crossover operator with an asexual crossover (PMX +A).

The rest of the paper is organized as follows: Section 2 presents the mathematical formulation of the MTSP. Section 3 reviews the different chromosome representations techniques reported so far the MTSP problem along with their pros and cons. Section 4 presents the framework of genetic algorithm with proposed group theory for population initialization and crossover. Section 5 presents the experimental results. Finally, Section 6 provides concluding remark of the study.

2 Formulations of Multiple Traveling Salesman Problem (MTSP)

In the MTSP, given a set of n nodes (cities) and m salesmen located at a single start (source), node. In this process the MTSP consists of finding tours for all m salesmen, who all start and end at the same source node, such that each city must be visited exactly once by only one salesman and the objective is to minimize the total cost of visiting all the nodes. This type of problem is modeled in graph theory as a weighted graph G = (V, E, w), where V is the set of vertices representing cities, E is the set of edges representing roads and w is weight (cost), between each pair of vertices.

A closed tour in which all the vertices are distinct which is known as Hamiltonian cycle. Finding set of m Hamiltonian cycles with minimum travel cost in the weighted graph gives the desired solution. i. e. total cost of visiting all nodes is minimized. The cost between cities is represented by matrix, known as cost matrix C = (cij), i, j = 1, 2, · · ·, n.

In cost matrix C the (i, j)th entry c ij, represents the cost of travel from i th to j th city. The matrix C is said to be symmetric when c ij=c ji, ∀ (i, j) ∈ E and asymmetric otherwise. In Integer Linear Programming framework the MTSP is formulated as follows [1]:

Let us define following decision variable xijk{0,1} as a binary variable whose value indicates whether a salesman visits next city or not. The value of variable xijk=1, if kth salesman selects an edge from city i to city j on the tour for travel, and xijk=0 if kth salesmen does not selects an edge between city i and city j on the tour for travel. Mathematically it can be expressed as follows:

xijk={1,ifkthsalesman travels from cityito cityj,0,otherwise. (1)

Consider Tk denote the tour traveled by k th salesman. The cost of the tour Tk travel by k th salesman is given as follows:

C(Tk)=i=1nj=1ncijxijk. (2)

The MTSP problem can be mathematically formulated as constraint optimization problem as follows:

minimizek=1mC(Tk), (3)

minimizemax(C(Tk)),for anyk, (4)

subject to

k=1mi=2nxi1k=m, (5)

k=1mi=2nx1jk=m, (6)

i=1nxijk=1,j=1,2,,n;ij;k=1,2,,m, (7)

j=1nxijk=1,i=1,2,,n;ij;k=1,2,,m, (8)

+subtour elimination constraints. (9)

Equation (3, 4) corresponds to the objective criterion i, ii of MTSP, as discussed in Section 1. The constraint given by Equation (7, 8) are usual city assignment constraint: at a time, a group of cities is assigned to exactly one salesman out of m salesmen. The constraint (5, 6), ensure that exactly m salesmen depart from and return back to home city (here it is represented by index 1). Constraint represented by Equation (9), is used to prevent sub-tours, which are degenerate tours that are formed between intermediate nodes and not connected to the home.

3 Chromosome Representations for the MTSP

The convergence and quality of solution obtained from GA is depends on the diversity in the search space. Thus, using an efficient encoding scheme for search space is necessary for success of GA. In general, in search space the solution / chromosomes are represented such that there is least chance to have multiple copy of the same solution i.e., search space have less redundancy and high diversity in the population. Here we present a brief review of different techniques for chromosome representation, which are commonly employed for solving MTSP, with their advantages and disadvantages. In proposed method, we used multi-chromosome representation technique, which is similar to the characteristic of the problem.

3.1 One Chromosome Technique

This method represents a solution for the MTSP using a single chromosome of length n + m − 1, where n represents a permutation (tour) of n cities with integer value ranging from 1 to n and m represents number of salesman. The solution chromosome is divided into m sub-tours by inserting m−1 dummy negative integers that represents the change from one salesman to another.

For example, in Fig.1 (where n = 10 and m = 3), the first salesman will visit cities 1, 3, and 6, the second salesman will visit cities 10, 5 and 9, and third salesman will visit 7, 4, 8 and 2. All visit in the order that the city appears. The number of possible solutions is (n + m - 1)!. However, many of the possible chromosomes are redundant. The detail of this technique is available in [5, 10, 31].

Fig. 1 One chromosome technique for a 10 city MTSP with three salesmen 

3.2 Two Chromosome Technique

This method represents a solution for the MTSP using a two chromosome of length n. one-chromosome represents a permutation (tour) of n cities with integer value ranging from integer 1 to n and the other chromosome represents city assigned to a salesman, and value ranging from integer 1 to m. For example, in Fig. 2, for n = 10 and m =3, the first salesman will visit cities 3, 10 and 2, the second salesman will visit cities 6, 5, 7 and 8, and third salesman will visit 1, 9 and 4. The visit is taken place in the same order as the cities appear in the permutation.

Fig. 2 Two chromosome technique for a 10 city MTSP with three salesmen 

In this method, there are n!mn possible solutions to the problem [6]. However, many of the possible solutions are redundant. For example, in Fig. 2, the first two genes in each of the chromosomes can be interchanged to create different chromosomes that result in an identical (or redundant) solution. For detail of two-chromosome technique one can refer [5, 6, 10, 31].

3.3 Two-Part Chromosome Technique

In this method, MTSP solution represented by two part of a chromosome of length n+m, where n is the number of city and m is the number of salesman used in MTSP. The first part of the chromosome of length n represents a permutation (tour), of n cities, in which each gene takes integer value ranging from 1 to n. Second part of chromosome of length m gives the number of cities assigned to each salesman.

The values assigned to the genes of the second part of the chromosome are constrained to be m positive integers, such that their sum is equal to the total number of cities to be visited. For example in Fig.3, the first salesman will visit cities 1, 5, 9 and 7, second salesman will visit cities 4, 8 and 6, and third salesman will visit 10, 3 and 2.

Fig. 3 Two parts chromosome technique for a 10 city MTSP with three salesmen 

In the two-part chromosome technique for the MTSP, there are n! possible permutations for the first part of the chromosome. The second part of the chromosome represents a positive vector of integers (k1,k2,,km), such that sum of their components must be equal to n.

There are (n1m1) distinct positive integer valued m-vectors that satisfy afore mentioned requirements [1]. Thus, the size of the solution space for the two-part chromosome representation is n!(n1m1). This is much lesser than the solution space size of the one chromosome, (n + m - 1)!, and two chromosome n!mn. The two-part chromosome technique reduces redundant solutions. The detail discussion about two part chromosome technique is given in [1, 5, 6].

3.4 Multi-Chromosome Technique

This method uses as many chromosomes as number of salesman [1]. The length of chromosome is variable and depends on the number of cities assigned to the particular salesman. Let the length of first chromosome is k 1 , the length of second chromosome is k 2 , and so on. Therefore, total number of cities in multi-chromosome representation equals to i=1mki=n. The assignment of the cities to the salesmen is random and in exclusive manner [1]. For example, in Fig. 4 (number of city n = 10 and number of salesman m = 3), the first salesman will visit cities 2, 5, 10 and 7, the second salesman will visit cities 9, 1 and 8, and third salesman will visit 6, 4, and 3.

Fig. 4 Multi-chromosome technique for a 10 city MTSP with three salesmen 

Determining number ways the cities assigned to the first chromosome is equal to the problem of obtaining an ordered subset of k 1 elements from a set of n elements. There are n!/(nk1)! distinct assignments. Since assignment is an exclusive, therefore for the second chromosome k 2 assignment is made out of nk1 cities. Therefore, we have (nk1)!/(nk1k2)! distinct assignment for second chromosome. Similarly, we can find the assignment of cities to other chromosomes. Therefore, total search space of the multi-chromosome techniques can be formulated as follows:

n!(nk1)!(nk1)!(nk1k2)!(nk1km1)!(nk1km)!=n!(nn)!=n! (10)

The length of the each chromosome represents a positive vector of integers (k1,k2,,km) that must sum to n.

There are (n1m1) distinct positive integer-valued m-vectors that satisfy this requirement [37].

Thus, size of the solution space for the multi-chromosome representation is n!(n1m1) it is equal to the two part chromosome representation, less than one chromosome and two chromosome representation technique. However, multi-chromosome representation is much similar to the characteristics of MTSP, salesmen are separated from each other “physically”, i.e., each salesmen visit different city except start and end city. This representation is more easily interpretable to the MTSP solution space.

Here we use multi chromosome representation [38] for encoding the solution / search space for the MTSP. The multi chromosomes representation has lesser number of duplicate solutions in search space, i.e., less redundancy, in comparison to other three chromosomes representation discussed so far [1, 5, 6, 10, 31].

4 The Genetic Algorithm for MTSP

We have developed a Genetic Algorithm for MTSP using a proposed group theory based method for tour construction and a new crossover operator. The proposed new crossover operator is inspired by the distance preserving crossover [11].

4.1 Group Theory for Tour Construction

There are various possible methods for generating the initial population [6, 23, 31, 34, 39, 40]. One of them is random generation of the initial population [6, 23, 31, 34], while another approach is to apply greedy constructive heuristic [39, 40] with Karp’s patching [20] to construct a feasible tour. Simplest and most straight forward method for generating the initial population is nearest neighbor tour construction heuristic [28].

But all aforesaid initialization methods suffer from inherent redundancy due to random initialization of the population. Singh et al. [31] try to overcome problem of redundancy, but it requires additional checking for uniqueness of newly generated chromosome against population members generated so far.

In this study, we generate the initial population using following iterative procedure. We assign label to the vertices of the weighted graph using the element of group of integers, Z n , with integer modulo n.

i+nj=(i+j)modn, (11)

where the addition on the right-hand side of equation is the ordinary addition of integers, the operation “+ n " is called addition modulo n. By using this function, we generate group table. In group table no two rows or columns are identical, it follows that every row of the composition table is obtained by a permutation of Z n and that each row / columns is a distinct permutation of n symbols [29]. The function P, which generates initial population, is defined as follows:

P(i)=(i+nj)+1, (12)

where i=1 to population size, and j=1 to n. In the group tour construction method, each individual in the initial population are unique that gives a wide diversity of genetic materials by exploring the whole search space. This due to initial heterogeneous population created using group theory that control diversity in generations and promote the exploration. Moreover, this method does not require any additional checking procedure for uniqueness of the population member.

4.2 Fitness

Our fitness function is same as the objective function. There are mainly two different objectives for the MTSP as discussed in introduction section and mathematically given by Equation (3) and (4), therefore, we have two different fitness functions. The objectives of MTSP are:

  • (i) Sum of the cost of m tours of the m salesman is minimum.

  • (ii) Maximum tour cost travel by any one salesman is minimized.

In both cases, we have to minimize the value of the fitness function.

4.3 Proposed Crossover Operator

The crossover operator is a method for sharing information between chromosomes. It combines the features of two parent chromosomes to form two offspring with the possibility that good chromosomes may generate better ones. Goldberg [13], described several order-based operators, such as the Partially Matched Crossover (PMX). The order crossover (OX) was suggested by Davis [9]. The position-based crossover (PBX) was introduced by Syswerda [33]. The cycle crossover (CX) was suggested by Oliver et al. [26]. Freisleben et al. [11] introduced a distance preserving crossover (DPX).Yuan et al. proposed a two-part chromosome crossover (TCX) [37].

In this paper, we propose a new crossover inspired by DPX crossover. In proposed crossover, the first city of chromosome s1 is copied to the last position on child c2 and last city of chromosome s1 is copied to the first position on child c2. Similarly, the first city of chromosome s2 is copied to the last position on child c1 and last city of chromosome s2 is copied to the first position on child c1.

The remaining cities changed as given algorithm in Fig. 6. For example, in Fig. 5, the first position of parents s1, s2 are copied to the last position in child c2 and c1 respectively and the last position of parent s1, s2 are copied to the first position in child c2 and c1 respectively, remaining positions 2, 3, 4, 5, 6, 7, and 8 cities are swapped as procedure given in Fig. 6.

Fig. 5 Proposed Crossover 

Fig. 6 Algorithm for proposed crossover 

The rationale behind this crossover is based on the intuitive idea that the city in sub optimal / optimal tours occurs at same position. In this crossover the Hamming distance is preserved. Upon use of this crossover there is very less chance to have duplicate members in the population.

4.4 2-Opt Optimal Mutation

Croes, G. A. [8] proposed 2-opt optimal mutation, and in literature it mostly named as 2-opt mutation. The 2-opt mutation replaces two edge from a tour by two new edges that are not in tour such that the cost of new tour is less than the original tour. This replacement process is continued till no further improvement in the cost of the new tour is possible, this is often referred to as 2-opt optimal. Here important to note that 2-opt mutation keeps the feasible tour corresponding to a reversal of a subsequence of the cities.

The method proceeds by replacing two non-adjacent edges (xi,xi+1) by two new edges (xi,xj) and (xi+1,xj+1), which are the only other two edges that can create a tour when the first two are deleted. In order to maintain a consistent orientation of the path by the predecessor-successor relationship, one of the two sub paths remaining after dropping the first two edges must be reversed [15].

For example, inverting the sub tour (xi+1,,xj) in tour (xixi+1,,xj,xj+1) using 2-opt mutation, we get the tour (xi,xj,,xi+1,xj+1). Finally, the change in the cost of the tour (xixi+1,,xj,xj+1) and (xi,xj,,xi+1,xj+1) is calculated as:

Δij=c(xi,xj)+c(xi+1,xj+1)c(xi,xi+1)c(xj,xj+1). (1)

The change in cost, Δ, gives the clue of the improvement of the tour. This process is repeated till Δ is negative. The generalized form of 2-opt process known as k-opt mutation is reported [16].

The 2-opt optimal mutation, which is sequential move and have chance of producing less diversify population for next generation i.e. increase in redundant solution in search space. The group theory tour construction algorithm for initial population generation has inherent capability of eliminating the redundant solutions.

Combination of 2-opt optimal mutation and Group theory tour construction algorithm, the exploration space for optimal tour gets reduced and hence the search time as well. Therefore, the combination of Group theory tour construction algorithm and 2-opt optimal mutation refines the tour for global optimality and decreases the time to get the optimal solution. We found that the proposed method gives better result in comparison to other heuristic methods reported in [37].

4.5 Proposed Algorithm

The first step of algorithm is population generation i.e. set of tour initialized using tour construction method described in Section 4.1. For multi-chromosome representation we randomly generate population of break-points of the tours for assigning the number of cities to each salesman for each tour. In the second step of algorithm, the fitness value is obtained for each tour in the population.

In the third step, randomly select ten tours and its corresponding break-points from the population. After then replace first tour and its corresponding break-points with minimum cost among selected ten tours and its break- points for crossover.

In the fourth step, apply proposed crossover operator on the first two tours from the ten tours obtained in step 4, with crossover probability rate (pc). After the crossover, the break points for both children are generated randomly.

In the last step, apply 2-opt optimal mutation operator on the tour having minimum cost among selected first two parents used for crossover in step four or new individuals that generated after crossover. After then update the population and its corresponding break-points. These processes repeated until specified generation is reached.

5 Experimental Results

5.1 Experimental Setup

For evaluating the performance of experimental results, Intel (R) Core (i5) 3.20 GHz processor, 2GB RAM on MATLAB is used. In this experiment, test problems are selected benchmark instances taken from the TSPLIB and reported in [6, 31, 37]. These test problems are Euclidean two-dimensional symmetric problem with 51, 100, and 150 cities. Here these problems are denoted as MTSP-51, MTSP-100, and MTSP-150. Number followed by MTSP indicates the number of city in the instance.

In Table 1, the experimental conditions of 12 different problem of sizes (n) and salesmen (m) combinations along with the number of generation used for each type of problem. The termination criterion is the number of generations. The value of parameters used in experiment are: population size (P) =100, tournament selection (k) =10, and crossover probability rate =0.85. The performance of methods compared based on best tour cost (Best), average tour cost (Mean), and standard-deviation (SD). Total 30 trails are made in order to collect the statistics of the result. A better method is considered to be those having lower value of Mean and Best than the other method.

Table 1 Experimental constraints 

Number of cities(n) Number of salesmen (m) GA generations
51 3, 5 and10 50000
100 3, 5, 10 and 20 100000
150 3, 5, 10, 20 and 30 200000

5.2 Experimental Results and Analysis

To evaluate the benefits of the proposed crossover operator and initial population generation technique, computational experiments were conducted and results are compared with the four crossover namely TCX, ORX +A, CYX +A and PMX +A reported in [35] on a set of MTSP problems. We conducted experiments considering both objective functions of MTSP.

The experimental result for first objective criterion - minimize the total travel distance of all the salesmen given by Equation (3) - is presented in Table 2 and graphically in Figure 7, Figure 8, and Figure 9. The cases where our proposed method obtains better result i.e. lower value of Mean and Best, than the all other method is shown in bold, while dash (-) values are not reported.

Table 2 Experimental Results for Minimisation of Total Travel Distance 

Instance Crossover m=3 m=5 m=10 m=20 m=30
Mean SD Best Mean SD Best Mean SD Best Mean SD Best Mean SD Best
MTSP-51 Proposed 466 6 460 515 10 499 693 20 669
TCX 510 24 466 536 26 499 636 17 602 - - - - - -
ORX+A 584 29 517 621 39 551 709 33 648 - - - - - -
CYX+A 591 43 511 622 44 530 710 42 633 - - - - - -
PMX+A 601 38 513 606 40 537 705 34 625 - - - - - -
MTSP-100 Proposed 24071 690 22959 26220 755 24559 35943 1221 33136 67623 2038 62963 - - -
TCX 32708 2267 28943 34179 2006 30941 36921 1964 32802 46976 1773 44112 - - -
ORX+A 41516 3356 36713 42716 2806 36196 44631 2997 38717 54265 3059 47971 - - -
CYX+A 41911 3195 35791 43634 2804 35421 45150 3241 40894 52916 2884 46466 - - -
PMX+A 41441 3423 33802 42063 3931 33908 44786 3467 39785 52142 2588 46212 - - -
MTSP-150 Proposed 40697 826 39504 42639 1825 39862 55895 1765 50892 79734 1138 77668 105072 922 102880
TCX 55851 2588 51126 61596 4759 51627 61360 3888 54473 69701 4340 62456 84008 5285 76481
ORX+A 67037 3745 60090 68018 3377 62539 72113 3637 63899 81696 5372 71933 96122 4562 88515
CYX+A 67463 4454 55335 69860 4342 61521 71584 4845 63126 83471 4197 75146 97106 3911 89008
PMX+A 68152 5140 58303 69112 4011 60761 72620 4334 64975 81178 4920 73281 95752 4923 87402

Fig. 7 Mean solution for the dataset MTSP-51 for different methods 

Fig. 8 Mean solution for the dataset MTSP-100 for different methods 

Fig. 9 Mean solution for the dataset MTSP-150 for different methods 

For all instance our proposed method gives better performance, with respect to all performance parameter for cases m = 3, 5. For m =10 our method gives slightly inferior solution in terms of Best for instance MTSP-51 and MTSP -100 of TCX, but it gives better average solution almost in all cases except TCX of MTSP -51. For m = 20, performance of proposed method degraded in terms of Best, but gives better Mean for instance MTSP-150 except the case TCX. For m = 30, the performance degraded with respect to all parameter.

The results for second objective criterion - minimizing the longest individual tour, which is also called make-span [37], given by Equation (4) - is presented in Table 3, and graphically in Figure 10-Figure 12. Clearly, proposed method outperforms all others method with respect to all parameters for all cases except TCX of MTSP- 100 for Best parameter.

Table 3 Experimental Results for Minimizing the Longest Tour 

Instance Crossover m=3 m=5 m=10 m=20 m=30
Mean SD Best Mean SD Best Mean SD Best Mean SD Best Mean SD Best
MTSP-51 Propose 188 4 184 139 3 129 112 0 112 - - - - - -
TCX 207 13 182 153 10 135 113 2 112 - - - - - -
ORX+A 216 12 191 165 11 139 128 13 112 - - - - - -
CYX+A 222 16 188 161 12 138 131 16 112 - - - - - -
PMX+A 218 11 191 161 10 141 130 12 112 - - - - - -
MTSP-100 Propose 10384 194 10031 7907 137 7728 6688 65 6581 6404 43 6363 - - -
TCX 14365 1013 12645 10086 674 8730 7768 492 6796 6768 433 6358 - - -
ORX+A 15137 1462 12997 11459 1053 9415 9286 1385 7373 8123 881 6666 - - -
CYX+A 15759 1242 13467 11333 1278 9507 9151 1364 7111 8109 936 6516 - - -
PMX+A 15238 1371 12249 11233 1177 9267 6890 1482 7187 8265 8608 6570 - - -
MTSP-150 Propose 15389 334 14804 13077 2539 10106 6884 99 6684 5546 36 5483 5251 5 5248
TCX 22523 1226 20556 16054 1227 14096 10722 927 8475 9640 789 8423 8759 806 7169
ORX+A 24766 1689 22015 17646 1519 15266 15150 2006 11788 13669 1816 10274 12204 1376 9182
CYX+A 23906 1941 20915 17608 1563 14029 14738 1750 10779 14111 1872 8365 13091 1295 10694
PMX+A 24216 1802 20347 17741 1235 15418 14489 2139 10738 13810 1315 11722 12608 1667 10080

Fig. 10 Maximum solution for the dataset MTSP-51 for different methods 

Fig. 11 Maximum solution for the dataset MTSP-100 for different methods 

Fig. 12 Maximum solution for the dataset MTSP-150 for different methods 

We observe that, as number of salesman (m) increases the performance of the proposed method with respect to first objective degraded. This is due to minimizing the longest tour balances the cities (or workload), among the salesmen and also minimizes the distance travelled by the salesmen. Minimization of the longest tour affects the fitness value as it gets decreased with the increasing number of salesmen.

6 Conclusion

In this paper, we developed a new crossover and initial population generation using group theoretic concept for solving multiple traveling salesperson problem (MTSP), using genetic algorithm (GA). Proposed method of initial population generation, gives the wide diversity in the population by ensuring uniqueness of the members in the population and thus redundancy due to random initialization is removed. Moreover, use of proposed crossover over this population has very less chance of producing redundant member in search space. We have compared the performance of the GA using proposed crossover and population initialization with the method using crossover TCX, ORX +A, CYX +A and PMX + A. Computational results reveals that the proposed technique outperform over all other method for second objective criterion, however, slightly inferior result for first objective in few cases (m =30 and m=20).


1 . Albayrak, M., & Allahverdi, N. (2011). Development a new mutation operator to solve the traveling salesman problem by aid of genetic algorithms. Expert Systems with Applications, Vol. 38, No. 3, pp. 1313-1320. [ Links ]

2 . Kiraly, A., & Abonyi, J. (2011). Optimization of MTSP by novel representation based on genetic algorithm. Intelligent Computing Optimization in Engneering, SCI 366, 241-269. [ Links ]

3 . Angel, R.D., Caudle, W.L., Noonan, R., & Whinston, A.N.D.A. (1972). Computer-assisted school bus scheduling. Management Science, Vol. 18, No. 6, B-279. [ Links ]

4 . Bektas, T (2006). The multiple traveling salesman problem: an overview of formulations and solution procedures. Omega, Vol. 34, No. 3, pp. 209-219. [ Links ]

5 . Brown, E.C., Ragsdale, C.T., & Carter, A.E. (2007). A grouping genetic algorithm for the multiple traveling salesperson problem. International Journal of Information Technology & Decision Making, Vol. 6, No. 2, pp. 333-347. [ Links ]

6 . Carter, A.E., & Ragsdale, C.T. (2006). A new approach to solving the multiple traveling salesperson problem using genetic algorithms. European journal of operational research, Vol. 175, No. 1, pp. 246-257. [ Links ]

7 . Chen, Y., & Zhang, P. (2006). Optimized annealing of traveling salesman problem from the nth-nearest-neighbor distribution. Physica A: Statistical Mechanics and its Applications, Vol. 371, No. 2, pp. 627-632. [ Links ]

8 . Croes, G.A (1958). A method for solving traveling-salesman problems. Operations research, Vol. 6, No. 6, pp. 791-812. [ Links ]

9 . Davis, L (1991). Handbook of Genetic Algorithms. Van Nostrand Reinhold New York Google Scholar. [ Links ]

10 . Falkenauer, E (1998). Genetic algorithms and grouping problems. John Wiley & Sons, Inc. [ Links ]

11 . Freisleben, B., & Merz, P. (1996). A genetic local search algorithm for solving symmetric and asymmetric traveling salesman problems. Evolutionary Computation Proceedings of IEEE International Conference, pp. 616-621. [ Links ]

12 . Gilbert, K.C., & Hofstra, R.B. (1992). A new multiperiod multiple traveling salesman problem with heuristic and application to a scheduling problem. Decision Sciences, Vol. 23, No. 1, pp. 250-259. [ Links ]

13 . Goldberg, E.D (1989). Genetic Algorithms in Search, Optimisation and Machine Learning. Adison Wesley PC. [ Links ]

14 . Gorenstein, S (1970). Printing press scheduling for multi-edition periodicals. Management Science, Vol. 16, No. 6, B-373. [ Links ]

15 . Gutin, G., & Punnen, A.P. (2006). The traveling salesman problem and its variations. Vol. 12, Springer Science & Business Media. [ Links ]

16 . Helsgaun, K (2000). An effective implementation of the Lin-Kernighan traveling salesman heuristic. European journal of operational research, Vol. 126, No. 1, pp. 106-130. [ Links ]

17 . Hao, J (2011). Preliminary Analysis on the Relative Solution Space Sizes for MTSP with Genetic Algorithm. Intelligent Computing and Information Science, pp. 649-655. [ Links ]

18 . Jeong, C.S., & Kim, M.H. (1991). Fast parallel simulated annealing for traveling salesman problem on SIMD machines with linear interconnections. Parallel computing, Vol. 17, No. 2-3, pp. 221-228. [ Links ]

19 . Johnson, D.S., & McGeoch, L.A. (1997). The traveling salesman problem: A case study in local optimization. Local search in combinatorial optimization, Vol. 1, pp. 215-310. [ Links ]

20 . Karp, R.M (1979). A Patching Algorithm for the Nonsymmetric Traveling Salesman Problem. SIAM J. Computing, Vol. 8, pp. 561-573. [ Links ]

21 . Liu, W., Li, S., Zhao, F., & Zheng, A. (2009). An ant colony optimization algorithm for the multiple traveling salesmen problem. 4th IEEE Conference Industrial Electronics and Applications, (ICIEA), pp. 1533-1537. [ Links ]

22 . Louis, S.J., & Li, G. (2000). Case injected genetic algorithms for traveling salesman problems. Information Sciences, Vol. 122, No. 2, pp. 201-225. [ Links ]

23 . Malmborg, C.J (1996). A genetic algorithm for service level based vehicle scheduling. European journal of operational research, Vol. 93, No. 1, pp. 121-134. [ Links ]

24 . Marinakis, Y., & Marinaki, M. (2010). A hybrid multi-swarm particle swarm optimization algorithm for the probabilistic traveling salesman problem. Computers & Operations Research, Vol. 37, No. 3, pp. 432-442. [ Links ]

25 . Okonjo-Adigwe, C (1988). An effective method of balancing the workload amongst salesmen. Omega, Vol. 16, No. 2, pp. 159-163. [ Links ]

26 . Oliver, I.M., Smith, D., & Holland, J.R. (1987). Study of permutation crossover operators on the traveling salesman problem. Proceedings of the second International Conference on Genetic Algorithms, Massachusetts Institute of Technology, NJ: L. Erlhaum Associates. [ Links ]

27 . Park, Y.B (2001). A hybrid genetic algorithm for the vehicle scheduling problem with due times and time deadlines. International Journal of Production Economics, Vol. 73, No. 2, pp. 175-188. [ Links ]

28 . Reinelt, G (1994). The traveling salesman: computational solutions for TSP applications. Springer-Verlag. [ Links ]

29 . Rosen, K.H., & Krithivasan, K. (2012). Discrete mathematics and its applications: with combinatorics and graph theory. Tata McGraw-Hill Education. [ Links ]

30 . Saleh, H.A., & Chelouah, R. (2004). The design of the global navigation satellite system surveying networks using genetic algorithms. Engineering Applications of Artificial Intelligence, Vol. 17, No. 1, pp. 111-122. [ Links ]

31 . Singh, A., & Baghel, A.S. (2009). A new grouping genetic algorithm approach to the multiple traveling salesperson problem. Soft Computing-A Fusion of Foundations, Methodologies and Applications, Vol. 13, No. 1, pp. 95-101. [ Links ]

32 . Song, C.H., Lee, K., & Lee, W.D. (2003). Extended simulated annealing for augmented TSP and multi-salesmen TSP. In Neural Networks. Proceedings of the International Joint Conference, Vol. 3, pp. 2340- 2343. [ Links ]

33 . Syswerda, G (1991). Schedule optimization using genetic algorithms. Handbook of genetic algorithms. [ Links ]

34 . Tang, L., Liu, J., Rong, A., & Yang, Z. (2000). A multiple traveling salesman problem model for hot rolling scheduling in Shanghai Baoshan Iron & Steel Complex. European journal of operational research, Vol. 124, No. 2, pp. 267-282. [ Links ]

35 . Venkatesh, P., & Singh, A. (2015). Two metaheuristic approaches for the multiple traveling salesperson problem. Applied Soft Computing, Vol. 26, pp. 74-89. [ Links ]

36 . Yousefikhoshbakht, M., Didehvar, F., & Rahmati, F. (2013). Modification of the ant colony optimization for solving the multiple traveling salesman problem. Romanian Journal of Information Science and Technology, Vol. 16, No. 1, pp. 65-80. [ Links ]

37 . Yuan, S., Skinner, B., Huang, S., & Liu, D. (2013). A new crossover approach for solving the multiple travelling salesmen problem using genetic algorithms. European journal of operational research, Vol. 228, No. 1, pp. 72-82. [ Links ]

38 . Yuan, S., Skinner, B.T., Huang, S., Liu, D.K., Dissanayake, G., Lau, H., & Pagac, D. (2011). A job grouping approach for planning container transfers at automated seaport container terminals. Advanced Engineering Informatics, Vol. 25, No. 3, pp. 413-426. [ Links ]

39 . Zhang, W (1993). Truncated Branch-and-Bound: A Case Study on the Asymmetric TSP. Proceedings of Spring Symposium on AI and NP-Hard Problems, pp. 160-166. [ Links ]

40 . Zhang, W (2000). Depth-First Branch-and-Bound versus Local Search: A Case Study. Proceedings of 17th National Conference on Artificial Intelligence (AAAI), pp. 930-935. [ Links ]

Received: August 09, 2017; Accepted: October 11, 2017

Corresponding author is Manoj Kumar Singh.,

Creative Commons License This is an open-access article distributed under the terms of the Creative Commons Attribution License