2021 International Conference on Computational Science and Computational Intelligence (CSCI)

2021 International Conference on Computational Science and Computational Intelligence (CSCI) | 978-1-6654-5841-2/21/$31.00 ©2021 IEEE | DOI: 10.1109/CSCI54926.2021.00148

Ant Colony Optimization for Balanced Multiple

Traveling Salesmen Problem
Bing Sun1, Chuan Wang2*, Qiang Yang3*, Weili Liu 4, and Weijie Yu 5

1. College of Computer and Information, Henan Normal University, Henan, China, [email protected]
2. College of Software, Henan Normal University, Henan Normal University, Henan, China, [email protected]
3. School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing, China, [email protected]
4. School of Computer Science, Guangdong Polytechnic Normal University, Guangdong, China, [email protected]
5. School of Information Management, Sun Yat-sen University, Guangdong, China, [email protected]

Abstract—Balanced multiple traveling salesmen problems

engineering, such as logistics distribution and transport
(BMTSP) are a popular kind of combinatorial optimization
problems [3], and vehicle routing problems [4].
problems widely existing in the real world. This problem aims
to minimize the total path length of all salesmen, and at the same
Compared with traditional TSP, MTSP and BMTSP are
time minimize the longest path among all salesmen to keep the more complex to optimize because in these two problems, not
path length balance. To solve this problem effectively, this paper only the allocation of cities to all salesmen need to be
proposes a balance biased ant colony optimization (BACO) optimized, but also the path of each salesman with respect to
algorithm. Specifically, this algorithm maintains ant groups to the allocated cities needs optimizing. Since BMTSP aims to
optimize the paths of all salesmen with each ant group optimize both the total path length of all salesmen and the
responsible for constructing a feasible solution and each ant in traveling cost balance between any two salesmen, the
a group responsible for building the path of one salesman. To optimization algorithms designed for traditional TSP cannot
construct balanced paths for all salesmen, this paper further effectively solve BMTSP.
develops four ant selection mechanisms to construct paths, In the literature, evolutionary algorithms (EAs), like
namely, Random Selection (RS), Shortest Biased Selection particle swarm optimization algorithms (PSO) [5-9], genetic
(SBS), Future Balance Biased Selection (FBBS) and Future algorithms (GA) [10], ant colony optimization algorithms
Shortest Biased Selection (FSBS). Additionally, we further (ACO) [11, 12], and estimation of distribution algorithms
introduce the 2-opt local search operation to optimize the path (EDA) [13], have shown very promising performance in
of each salesman. Finally, extensive experiments conducted on solving complicated optimization problems. Particularly,
four TSPLIB benchmark sets with different numbers of confronted with discrete optimization problems, like TSP,
salesmen demonstrate that the proposed BACO with the four
ACO has presented great superiority to other EAs [14, 15].
ant selection mechanisms shows much better performance than
Taking advantage of this merit of ACO, many researchers
a state-of-the-art genetic algorithm (GA). In particular, among
the four selection mechanisms, the FSBS strategy helps BACO
have been devoted to designing novel ACO to solve
achieve the best performance in solving BMSTP. complicated MTSP and BMTSP. To name a few, Wang et al.
[16] proposed a new pheromone update rule for ACO to
Keywords- Ant Colony Optimization; Multiple Traveling improve its search efficiency in solving MTSP. Harrath et al.
Salesmen Problem; Balance Biased Solution Construction; [17] combined ACO with GA and the 2-opt algorithm to
Travelling Salesman Problem; Local Search generate feasible and promising solutions for MTSP. Qi et al.
[18] incorporated the simulated annealing mechanism into
I. INTRODUCTION ACO to accelerate the convergence of the ant colony.
Although these ACO variants have achieved promising results
Balanced multiple traveling salesmen problem (BMTSP) in optimizing the total path length of all salesmen, they do not
is an extension of the multiple traveling salesmen problem consider the traveling cost balance between any two salesmen.
(MTSP), and both are NP hard problems [1]. In MTSP, several To fill the above gap, researchers have devised novel ACO
salesmen start from their departure cities and cooperate to by taking the traveling cost balance among salesmen into
complete the traversal of other cities. Similar to TSP, in MTSP, consideration. For instance, Chen et al. [12] proposed an
each city except for the start cities must be visited only once. approach to divide the ant colony into several groups and
The objective of MTSP is to minimize the total cost of all maintain a separate pheromone matrix for each group. Then,
salesmen. Besides the above objective in MSTP, another each group of ants is utilized to optimize one objective. At the
objective to keep the cost balance between any two salesmen end, a set of Pareto solutions is obtained by taking the optimal
is taken into consideration in BMTSP [2]. In particular, it is solution from each group. Lu et al. [14] incorporated a max-
found that BMTSP is very common in daily life and industrial min search strategy into ACO by giving priority to the ant with
This work was supported in part by the National Natural Science Foundation of the minimal path length, so that ACO can guarantee the
China under Grant 62006124 and 61873097, in part by the Natural Science traveling cost balance in the path building process of ants.
Foundation of Jiangsu Province under Project BK20200811, in part by the Natural
Science Foundation of the Jiangsu Higher Education Institutions of China under
Lupoaie et al. [19] proposed a cross-tour mutation for ACO to
Grant 20KJB520006, and in part by the Startup Foundation for Introducing Talent let long paths crossover with short tour paths with high
of NUIST (*Corresponding authors: Chuan Wang and Qiang Yang). probability to balance the traveling cost between salesmen.

978-1-6654-5841-2/21/$31.00 ©2021 IEEE 476

DOI 10.1109/CSCI54926.2021.00148
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PIAUI. Downloaded on December 30,2024 at 14:25:06 UTC from IEEE Xplore. Restrictions apply.
 Though the above methods could solve BMTSP to some Pathi Path j ‡ i z j ф1 d i , j d M
extent, they do not consider the priority of ants, when the paths
of ants are unbalanced during the path building process [12]. M (1)
To compensate for this shortcoming, this paper proposes a ¦ N N
¯i 1 i
balance biased ACO (BACO) for BMTSP. Secondly, let the path length of the ith salesman be Ci, and
Specifically, BACO maintains one ant to build the path of then the total length L of all paths can be calculated as
each salesman as in [14]. Therefore, for each feasible solution,
¦
M
an ant group, whose size is equal to the number of salesmen, L i 1
Ci (2)
is employed to construct the paths of all salesmen. In Eventually, the objectives of BMTSP can be defined as
particular, these ants build the paths in parallel by selecting
cities sequentially. Since one city is exclusively assigned to Min L
(3)
one salesman during the path building process, only one ant in Min Max{Ci , 1 d i d M }
the group is allowed to select a city from the rest cities once a
To solve BMTSP effectively, optimization algorithms
time. To guarantee the traveling cost balance among salesmen,
we develop four ant selection mechanisms to select one ant should search the solution space by optimizing the above two
from the group to go to the next city. These four selection objectives simultaneously.
mechanisms are: 1) Random Selection (RS), which randomly III. ANT COLONY OPTIMIZATION
selects an ant from the group to select the next city; 2) Shortest
Biased Selection (SBS), which selects the ant with the shortest The basic idea of ACO was inspired by the foraging
path to further build the path; 3) Future Balance Biased process of ants. Specifically, the optimization process of ACO
Selection (FBBS), which first selects the ant with the shortest is elucidated as follows. First, in the initial stage, since no city
distance to the rest cities if its path is not the longest in the ant has been visited, the pheromone deposited on each edge is
colony, otherwise selects the ant with the shortest path length; initialized by a same value. Then, during the path building
and 4) Future Shortest Biased Selection (FBSS), which also process of each ant k (k=1,2,.., pop_size, pop_size is the
first selects the ant with the shortest distance to the rest cities number of ants), the cities already visited by this ant are added
if its path is not the longest, otherwise selects the ant with the to a tabu list tabuk. Given that ant k is at city i, it chooses the
second smallest distance to the rest cities. next city in its path from the remaining cities based on the
To verify the effectiveness of the proposed BACO, we roulette wheel selection with the selection probability of each
first select 4 benchmark TSP sets from the popular TSPLIB city computed as follows:
and then generate 3 different numbers of salesmen for each ­ [ τij (t )]α [ηij (t )]β
instance. Subsequently, extensive experiments are conducted ° , if j  J k (i )
® ¦ sJ ( i ) [ τ is (t )] [ηis (t )] (4)
k α β
on these generated instances by comparing the proposed Pij
°
k

BACO with a state-of-the-art GA method. Additionally, deep

investigation on the four selection mechanisms is also taken ¯0, otherwise.
to observe their influence on BACO in solving BMTSP. where Jk(i) = E-tabuk, E is the set of all cities, and tabuk stores
The rest of this paper is organized as follows. Section II the cities that ant k has already visited. τij(t) is the pheromone
illustrates the mathematical definition of BMTSP. Section III on the edge connecting city i and city j, while ηij(t) is the
introduces the basic ACO for TSP. In Section IV, we heuristic information, which is usually set as the reciprocal of
introduce the proposed BACO and the proposed four selection the distance between city i and city j. α is a parameter
mechanisms. In Section V, we conduct experiments to verify controlling the influence of the pheromone, while β is a
the effectiveness of the proposed BACO. Finally, Section VI parameter in charge of the effect of the heuristic information.
concludes this paper. After all ants finish constructing their paths, the traveling
cost of each ant is evaluated. Then, the pheromone on each
II. BALANCED MTSP edge is updated as follows:
In BMTSP, multiple traveling salesmen start from their W ij (t  1) (1  U )W ij (t )+'W ij (t ) (5)
own depots and then cooperate to visit all cities with each city
pop _ size
visited exactly once by only one salesman. Two objectives are
involved in BMTSP, namely, minimizing the total path length
'τ ij (t ) ¦
k 1
'τijk (t ) (6)
of all salesmen and keeping the path length balance among all
salesmen. As for the second objective, this paper realizes it by ­ Q , if ant k passes city i to city j
k ° (7)
minimizing the longest path among all salesmen. 'τ ij (t ) ® Lk
For better understanding, we give the mathematical °0, otherwise
definition of BMTSP as follows. Firstly, suppose M salesmen
¯
where Lk is the path length of the kth ant, pop_size is the
and N cities (M<N) are involved in BMSTP. Let Pathi
represent the path of the ith salesman and Ni be the number of number of ants, ρ is the pheromone evaporation rate, and Q
cities in Pathi. According to the requirements that all paths is a constant parameter. Uτij(t) is the sum of the pheromone
cannot intersect with each other and all cities must be visited deposited by all ants on the edge between city i and city j.
exactly once only by one salesman, the following constraints Overall, the general pseudocode of ACO is outlined in
should be satisfied: Algorithm 1. Specifically, at the beginning, as shown in Line

477

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PIAUI. Downloaded on December 30,2024 at 14:25:06 UTC from IEEE Xplore. Restrictions apply.
 Algorithm 1: The pseudocode of ACO which may even have different orders of magnitude,
Input: maximum number of iterations FESmax, number of cities N, number especially on large-scale BMTSP either with a lot of cities or
of ant groups pop_size. with a large number of salesmen. After normalization, the
1: Initialize the pheromone matrix with a constant value, and set count= 0;
2: While (countō FESmax) do two values are limited within (0, 1), and thus the difference
3: For i=1 to pop_size is narrowed. In this way, the two objectives make the same
4: Set tabu=ѫ;
5: Randomly select a start city for ant i and add this city to tabu; influence on the optimization of BMTSP, and thus there is
6: While (the size of tabu !=N) likely no bias in the final optimization results.
7: Calculate the selection probability by Eq. (4);
8: Ant i chooses the jth city to visit by the roulette wheel selection; In this paper, we employ Eq. (8) to evaluate the quality
9: Add the selected city to tabu; of each feasible solution, so that both minimizing the total
10: End While
11: Calculate the path length of the ith ant; path length of all salesmen and keep the path length balance
12: End For among all salesmen are taken into consideration.
13: Update the pheromone matrix based on Eq. (5) and count++;
14:End While B. Path Building
15:Obtain the global best solution gbest and its fitness f(gbest);
Output: f(gbest) and gbest Similar to [12], this paper maintains a population of ant
groups in BACO to optimize BMTSP. Specifically, each ant
1, the pheromone matrix initialized with a constant value. group contains M ants, where M is the number of salesmen in
Then, it goes to the main iteration (Lines 2-14) of the BMTSP. In the ant group, each ant is responsible for building
algorithm. First, each ant constructs its path city by city as the path of one salesman. During the path building process,
shown in Lines 4-10. Specifically, each ant first randomly different from traditional ACO for TSP, the M ants in one
selects a city as the start city and this city is added into the group construct the paths of the M salesmen in parallel.
tabu list as shown In Line 5. Then, it selects the next city to Specifically, the paths of the M ants in one group are first
move based on the roulette wheel selection with the initialized as follows:
probability computed by Eq. (4). Once a city is chosen, it is 1) Randomly assign a unique start city to each ant and add
added into the tabu list. After the ant finish building its paths, these start cities into the tabu list, and store the M ant
it is evaluated as shown in Line 11. Subsequently, the into an ant list;
pheromone matrix is updated by Eq. (5) in Line 13. After the 2) Each ant in the ant list pre-selects the next city from the
termination condition is met, the algorithm outputs the found remaining cities through the roulette wheel selection
global best solution. with the probability computed based on Eq. (4);
IV. BALANCE ANTD COLONY ALGORITHM 3) Select the ant with the minimal distance between the
start city and the selected city to build the path; Then,
To effectively solve the BMTSP, we propose a balance add this city to the tabu list and exclude the selected ant
biased ant colony optimization algorithm (BACO) in this from the ant list;
paper by taking the current path lengths of ants into 4) Repeat Step 2, and Step 3, until the ant list is empty. In
consideration during the path building process. The details of this way, each ant has two cities in its current path.
this algorithm are elucidated in the following. The above initialization process of an ant group
A. Definition of Fitness Function (containing M ants) is to guarantee that each salesman could
visit at least one city except for its start city.
This paper transforms the two objectives in Eq. (3) into
After the initialization process, the remaining paths of
a single objective as follows:
the M ants are constructed city by city. However, different
minF Li '  MaxCi ' (8)
from the path building process of ACO for traditional TSP,
Li  Lmin in the proposed BACO, M ants construct their paths in
Li ' (9)
Lmax  Lmin parallel. Nevertheless, to guarantee the feasibility of the
solution constructed by the M ants, all ants select the next
MaxCi  MaxCmin (10)
MaxCi ' target city in their paths sequentially, so that a city can only
MaxCmax  MaxCmin appear in only one path. Due to this, there is a priority issue
where Lmax and Lmin are the maximum and the minimum of among the M ants, namely which ant should be selected to
the total path length among all feasible solutions, respectively; construct the path first.
L'i is the normalized total path length of the ith solution To solve the above priority issue, this paper devises the
computed by Eq. (9). Similarly, MaxCmax and MaxCmin are the following four different ant selection mechanisms:
maximum and the minimum of the longest path length (MaxC) (1) Random Selection (RS): This strategy randomly selects
among all salesmen in all feasible solutions, respectively. an ant from the M ants with equal probability to build the
MaxC'i is the normalized longest path length among all path. Such a strategy is very simple, but it does not take
salesmen in the ith solution. the path length balance into consideration. Therefore, it
It should be mentioned that the two objective values of is very likely that some ants may preserve very long
each feasible solution is normalized here because the original paths. This method is used as a baseline method to
values of the two objectives may preserve large difference, compare with other methods.

478

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PIAUI. Downloaded on December 30,2024 at 14:25:06 UTC from IEEE Xplore. Restrictions apply.
 Ant1 length:100 6 11 7 distance from 7 to 12 is 30
12 Algorithm 3: The pseudocode of BACO
distance from 4 to 13 is 10
Input: maximum number of iterations FESmax, number of cities N, number
Ant2 length:130 3 5 8 4 13 of salesmen M, number of ant groups pop_size.
distance from 10 to 1 is 20
1: Initialize pheromone matrix with a constant value, and set count= 0;
Ant3 length:120 2 9 10 1 2: While (countō FESmax) do
Fig 1. Examples of the ant selection 3: For i=1 to pop_size
4: Set tabu=ѫ;
(2) Shortest Biased Selection (SBS): This strategy first 5: Randomly select a unique city as the start city for each ant in the
6: ith ant group and add these M cities to tabu;
compares the current path length of all ants, and then Each ant selects the next city nearest to its start city sequentially,
selects the ant with the shortest path to build the path. 7: and add these M cities to tabu;
Utilize one of the four developed ant selection strategies to select
For example, as shown in Fig 1, in the current situation, 8: ants to build the paths based on Algorithm 2;
Ant 1 has the shortest path of 100 among all the three 9: Conduct the 2-opt method on each ant to optimize the path;
10: Calculate the fitness of the constructed solution based on Eq. (8);
ants. Therefore, Ant 1 is selected to build the path. This 11: End For
strategy is always biased to the ant with the shortest path 12: Update the pheromone matrix based on Eq. (5), and count++;
13:End While
in the current situation, and thus is a greedy strategy. It 14:Obtain the global best solution gbest and its fitness f(gbest);
can maintain the path length balance among all salesmen Output: f(gbest) and gbest
to a certain degree.
strategy is very similar to FBBS. The only difference is
(3) Future Balance Biased Selection (FBBS): In this
that when the current path length of the ant with the
strategy, all ants first pre-select a city from the remaining
minimal distance between the selected city and the last
cities based on the roulette wheel selection with the
city in the path is the longest, this strategy selects the ant
probability computed based on Eq. (4). Then, for each
with the second minimal distance between the selected
ant, the distance between the selected city and the last
city and the last city in the path. For example, in Fig. 1,
city in its path is computed. The ant with the minimal
Ant 2 is first selected because the distance between the
distance is selected to build its path. However, if the
selected city 13 and its last city 4 in the path is the
current path length of the selected ant is the longest, this
smallest. However, the current path length of this ant is
strategy will select the ant with the shortest current path
the largest. Therefore, this strategy selects Ant 3 to build
length to build the path. For example, in Fig. 1, the three
the path because the distance between the selected city 1
ants select a city from the remaining cities and the
and its last city 10 in the path is the second smallest.
distance between the selected city and the last city in the
With the above four ant selection strategies, an ant group
path of each ant is computed. Among them, the distance
could construct reasonable paths for all salesmen and
from city 4 to city 13 is the shortest for Ant 2. Therefore,
eventually a feasible solution is formed. The path building
Ant 2 is selected to build the path. However, it is found
process with the four strategies is outlined in Algorithm 2.
that the current path length of Ant 2 is the longest. If we
further select this ant to build the path, the unbalance C. Overall Framework of BACO
among the path length will be enlarged. Hence, to keep Integrated the above techniques into ACO, the developed
the path length balance, this strategy selects the ant with BACO is outlined in Algorithm 3. Specifically, similar to the
the shortest path length, namely Ant 1, to build the path. traditional ACO, BACO first initializes the pheromone
(4) Future Shortest Biased Selection (FSBS): This matrix with a same constant value. Then, the algorithm goes
Algorithm 2: The pseudocode of Path Building to the main loop to search the solution space iteratively.
Input: ant selection strategy ASS; number of cities N, an initialized ant In the main loop, during each generation, each ant group
group; tabu list tabu with M ants generates a feasible solution by constructing the
1: While( the size of tabu!= N)
2: Calculate the selection probability by Eq. (4); paths of all salesmen based on Algorithm 2. Then, the 2-opt
3: Each ant pre-chooses a city to visit by the roulette selection; local search method is conducted on the built paths of all
4: Switch (ASS)
5: Case: RS salesmen to further optimize the paths in Line 9. After that,
6: Randomly select an ant to build the associated path; the fitness of the generated solution is evaluated based on Eq.
7: Case: SBS (8). After all ant groups finish constructing the solutions, the
8: Select the ant with the minimal path length to build the path;
9: Case: FBSS pheromone matrix is updated based on Eq. (5). After the main
10: Pre-select the ant with the minimal distance between the last city loop terminates the global best solution is obtained and output.
in the path and the selected city;
11: If the path of this ant is the longest It should be mentioned that the 2-opt local search method
12: Select the ant with the minimal path length to build the path; is adopted to optimize the constructed path of each ant in each
13: End If
14: Case: FSBS ant group. This is because ACO is lack of subtle refinement
15: Pre-select the ant with the minimal distance between the last city to improve the solution accuracy. Such a technique widely
in the path and the selected city;
16: If the path of this ant is the longest exists in the literature [20] to help ACO optimize various
17: Select the ant with the second minimal distance between the last optimization problems with high accuracy.
city in the path and the selected city;;
18: End If V. EXPERIMENTS
19: Let the selected ant to move to the selected city and add it into tabu;
20: End While This section verifies the effectiveness of the proposed
Output: The built paths of all salesmen (namely a feasible solution)

479

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PIAUI. Downloaded on December 30,2024 at 14:25:06 UTC from IEEE Xplore. Restrictions apply.
 Table I. Optimization Results of BACO with the Four Selection Mechanisms on the Four TSP Sets with Different Numbers of Salesmen
BACO-RS BACO-SBS
No.sm Test set Best Worst Mean Best Worst Mean
Total Longest Total Longest Total Longest Total Longest Total Longest Total Longest
Att48 34733 12933 39242 13534 37169 13032 34244 11855 38018 12836 36075 12448
kroA100 23532 8434 27012 9176 25187 8813 23068 7811 25810 8715 24604 8401
3
kroA150 30661 10654 34037 11616 32628 11293 30211 10277 32702 11131 31422 10698
a280 3158 1069 3497 1186 3352 1154 3124 1050 3429 1151 3307 1121
Att48 36232 10283 42319 11084 38884 10751 35848 9427 38754 9855 37033 9575
kroA100 26173 6746 29472 7650 27535 7299 24076 6153 26508 6804 25422 6556
4
kroA150 33144 8605 36476 9556 34861 9214 31544 8264 35479 9114 33739 8692
a280 3422 875 3687 948 3496 925 3281 853 3610 912 3454 885
Att48 36962 8724 46590 9832 40827 9311 36704 7614 41728 8696 39091 8299
kroA100 26401 5772 30758 6606 28579 6276 25882 5323 30407 6175 28234 5904
5
kroA150 33332 7348 39139 8450 36742 7967 32592 6724 37217 7668 35385 7357
a280 3427 708 3850 811 3660 784 3408 700 3780 774 3597 745
BACO-FBBS BACO-FSBS
No.sm Test set Best Worst Mean Best Worst Mean
Total Longest Total Longest Total Longest Total Longest Total Longest Total Longest
Att48 34103 11675 36655 12622 35099 12154 33810 11471 35273 12215 34539 11926
kroA100 22404 7684 25235 8457 24196 8208 22415 7688 24786 8376 23855 8132
3
kroA150 30130 10232 33389 11228 32144 10895 29317 10149 31947 10804 30897 10507
a280 3100 1046 3401 1141 3253 1097 2997 1019 3185 1095 3121 1059
Att48 34851 9107 38095 9762 36530 9484 34691 9319 37472 9519 36462 9513
kroA100 23676 6171 26440 6798 24995 6486 25196 6433 28484 7224 24774 6496
4
kroA150 30600 8047 33880 8707 32472 8387 29354 7416 33579 8558 31734 8327
a280 3150 810 3465 879 3331 856 3058 801 3378 854 3224 833
Att48 35929 7395 39051 8049 36590 7701 35069 7348 38434 8123 36497 7674
kroA100 24351 4999 28009 5807 25962 5441 22638 5178 26944 5592 25287 5381
5
kroA150 30883 6452 34583 7249 33110 6965 30153 6462 33567 7059 32010 6815
a280 3298 693 3533 730 3411 709 3173 673 3297 688 3218 680

BACO by conducting various experiments. First, we select From Table I, we find that: 1) BACO-FSBS gets the best
four popular TSP sets of different sizes from the popular performance, while BACO-RS gets the worst performance
TSPLIB, namely Att48, kroA100, kroA150, and a280. Then, among all BACO with the four ant selection strategies. 2) The
for each set, we assign three different numbers of salesmen, averaged longest path length of BACO-FSBS is much
namely 3, 4, and 5. Therefore, a total number of 12 BMTSP smaller than that of the other three BACOs. 3) Next to
instances are generated. BACO-FSBS, BACO-FBBS achieves the second best
To make comparison with the proposed BACO, we performance among all the BACO variants.
utilize an improved genetic algorithm (GA) [21] as the The above observations indicate that FSBS and FBBS
compared method. To make fair comparison, we also utilize are much superior to RS and SBS in help ACO to effectively
the fitness function in Eq. (8) to evaluate the fitness of solve BMTSP. This is because RS is a completely random
individuals in GA. In addition, the 2-opt local search method strategy, which does not take the path length balance into
is also incorporated into GA to improve solution accuracy. consideration. As a result, BACO-RS achieves the worst
As for the parameter settings, for the proposed BACO, performance. SBS is a greedy selection strategy based on the
the number of ant groups is set as 30, the control parameters current situation. The decision with respect to the ant
α=1, and β=5, the constant parameter Q=1, and the selection is only suitable for the current state. It does not
evaporation rate ρ=0.1. As for the compared GA, the consider the future state. Therefore, BACO-SBS also
crossover rate is 0.4, the mutation rate is 0.1, and the number achieves very poor performance. Regarding FBBS and FSBS,
of chromosomes is 30. For fairness, the number of iterations they both make the decision based on the future state. Besides,
is 1000 for all algorithms. they not only optimize the total path length as much as
First, we investigate the influence of the developed four possible, but also prevent the path length imbalance by
ant selection strategies on the proposed BACO. Table I shows avoiding excessive extension of the longest path. Therefore,
the comparison results among different BACO on each both BACO-FSBS and BACO-FBBS achieve very promising
instance over 30 independent runs. In this table, the total path performance in solving BMTSP.
length (denoted as “Total”) and the longest path length Then, we compare the four different BACO variants with
(denoted as “longest”) of the best solution (denoted as the compared GA. Fig. 2 displays the comparison results
“Best”), and the worst solution (denoted as “Worst”) are between the four BACO variants and GA on the four TSP sets
presented. In addition, the averaged total path length and the with 4 salesmen.
averaged longest path length are also exhibited in this table. From this figure, we find that given the same number of
00 0000 0000 000
00 5000 500
27535
45758

27115

5000
3533
3496
34861

00
3454
25422

33739

33478
24995

3331
24774

0000 000
32472

3224
31734
38884

00
37033

0000
36530

36462

00 5000 500

00 5000 0000 000

00 5000 500
0000
00
0000 000
00
7299

972
11833

5000
925
9214
6797

885
6556

8747
8692
6496
6486
10751

856
8387

8327

833

5000 500
9575

9513
9484

00
0 0 0 0
RS SF FBBS FSBS GA RS SF FBBS FSBS GA RS SF FBBS FSBS GA RS SF FBBS FSBS GA
Longest Total Longest Total Longest Total Longest Total
(a).Results of att48 (b).Results of kroA100 (c).Results of kroA150 (d).Results of a280

Fig 2. Comparison results between BACO with different ant selection schemes and GA on the 4 TSP sets with 4 salesmen.

480

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PIAUI. Downloaded on December 30,2024 at 14:25:06 UTC from IEEE Xplore. Restrictions apply.
 sm1 sm2 sm3 sm4 sm1 sm2 sm3 sm4 sm1 sm2 sm3 sm4 sm1 sm2 sm3 sm4 sm1 sm2 sm3 sm4
2000 16
14 34 2000 16
14 34 2000 16
14 34 2000 16
14 34 2000 34
16
14
58 10 44 8526 42 98 58 10 44 8526 42 58 10 44 8526 42 58 10 44 8526
85 42 58
98 59 61 59 61 98 59 61 98 59 61 98 10 44 59 61 85 26 42
73 40 73 40 73 40 73 40 73
31 22 20 31 22 31 22 20 31 22 31 40
35
20 19 11
45 2
70 35 9097 19 11 45 2 70 35
20 19 11
45 2
70 35 90
9097 19 11 45 2 70 20 22
19 11 70
90 97
76 54 76 54
90 97
76 54 76 54
35 90 97 45 2
82 82 82 82 76 54
82
71 99 71 33 13 99 71 99 71 33 13 99 71 99
1500 37
56 6
33
28
13
1500 37 566 28 1500 37
566
33
28
13
1500 37 566 28 1500 37
56 6
33
28
13
23 8 23 8 23 8 23 8 23 8
83 46 83 46 83 46 83 46 83 46
9 47 9 47 9 47 9 47 9 47
86 86 86 86 86
17 17 17 17 17
50 50 50 50 50

78 62
29
78 62 29 78 62
29
78 62 29 29
1000 89
1000 89 92 1000 89
1000 89 92 78 62
Y

Y
92 60 60 92 60 60 1000 89

Y
93 5
48 27 57 93 5
48 27 57 93 5
48 27 57 93 5
48 27 57 93 92
57 60
0 24 51 0 24 51 0 24 51 0 24 51 485 0 27 24
87 80 84 9577 87 80 84 9577 87 80 84 9577 87 80 84 9577 87 80 51
2115 2115 2115 2115 2115
84 9577
52
67
38 52 67 38 52
67
38 52 67 38 52 38
67
69
91 66 69 18 91 66 91 66 69 18 91 66 91
18 4 4 69 18 4 4 69 18 66 4
74 68 36 74 68 36 74 68 36 74 68 36 74 68 36
72
7 72 72
7 72 72
500 7
49 500 49 500 7
49 500 49 500 7
49
96 96 96 96 96
65 3 43 65 3 43 65 3 43 65 3 43 65 3 43
41 63
55 41 63 41 63
55 41 63 63
64 55
39 64 39 64 55
39 64 39 64 55 41
39
81 12 32
53 81 1232 81 12 32
53 81 1232 12 32
88 53
79 88 75
88 53
79 88 75 88 53 81
79
30
1 75
30 1 79
30
1 75
30 1 79 1 75
94 94 94 94 30
0 25
0 25 0 25
0 25
0 25 94

0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000
X X X X X

(a).GA (b).BACO-RS (c).BACO-SBS (d).BACO-FBBS (e).BACO-FSBS

Fig 3. Optimal paths of the 4 salesmen obtained by GA and different BACO variants on the kroA100 TSP set.
iterations, the four BACO variants all achieves much better [2] R. S. Beed, S. Sarkar, A. Roy, and S. Chatterjee, "A Study of the
Genetic Algorithm Parameters for Solving Multi-objective Travelling
performance than GA in both the longest path length and the
Salesman Problem," in Int. Conf. Inf. Technol., 2017, pp. 23-29.
total path length. This verifies that the four ant selection [3] D. Krstev et al., "The Multiple Travelling Salesman Problem and
strategies are very beneficial for ACO to solve BMTSP. Vehicle Routing Problem for Different Domestic Drinks," Int. J. Sci.
To further observe the difference between BACOs and Tech. Innov. Ind. MTM, vol. 22, pp. 96-98, 2014.
GA, we plot the paths of all salesmen obtained by each [4] N. Zakerinejad, D. Riera, and D. Guimarans, "ACO and CP Working
Together to Build a Flexible Tool for the VRP," in Int. Conf. Oper. Res.
algorithm when it terminates. Fig. 3 presents the optimal Enterp. Syst., 2016, pp. 122-129.
paths of all salesmen in the best run of all algorithms on the [5] Q. Yang, W. N. Chen, T. Gu, H. Jin, W. Mao, and J. Zhang, "An
kroA100 test set with 4 salesmen. Adaptive Stochastic Dominant Learning Swarm Optimizer for High-
From this figure, we find that the paths of all salesmen is Dimensional Optimization," IEEE Trans. Cybern., pp. 1-17, 2020.
[6] F. F. Wei et al., "A Classifier-Assisted Level-Based Learning Swarm
optimal under the allocated cities. However, taking deep Optimizer for Expensive Optimization," IEEE Trans. Evol. Comput.,
observation, we find that the allocation of cities to the vol. 25, pp. 219-233, 2021.
salesmen is very different. This might be the main reason that [7] Q. Yang et al., "A Distributed Swarm Optimizer With Adaptive
Communication for Large-Scale Optimization," IEEE Trans. Cybern.,
leads to the optimization difference. Given the allocation of
vol. 50, pp. 3393-3408, 2020.
cities in BACO-FSBS is the optimal since it achieves the best [8] Q. Yang, W. Chen, J. D. Deng, Y. Li, T. Gu, and J. Zhang, "A Level-
performance, we find that the allocation of cities in GA, Based Learning Swarm Optimizer for Large-Scale Optimization,"
BACO-RS, and BACO-SBS are very different from the IEEE Trans. Evol. Comput. , vol. 22, pp. 578-594, 2018.
optimal one. This indicates that the allocation of cities is very [9] Q. Yang et al., "Segment-Based Predominant Learning Swarm
Optimizer for Large-Scale Optimization," IEEE Trans. Cybern., vol.
vital in BMTSP and the proposed FSBS and FBBS could 47, pp. 2896-2910, 2017.
allocate cities to the salesmen very effectively. [10] R. Saxena, M. Jain, S. Bhadri, and S. Khemka, "Parallelizing GA
To summarize, based on the above experiments, we can Based Heuristic Approach for TSP over CUDA and OPENMP," in Int.
see that BACO variants are very effective to solve BMTSP, Conf. Adv. Comput., Commun. Inf., 2017, pp. 1934-1940.
[11] Q. Yang et al., "Adaptive Multimodal Continuous Ant Colony
especially the ones with the FSBS and the FBBS. Optimization," IEEE Trans. Evol. Comput., vol. 21, pp. 191-205, 2017.
[12] X. Chen et al., "Ant Colony Optimization Based Memetic Algorithm
VI. CONCLUSION to Solve Bi-objective Multiple Traveling Salesmen Problem for Multi-
Robot Systems," IEEE Access, vol. 6, pp. 21745-21757, 2018.
This paper has proposed a balance biased ant colony [13] Q. Yang et al., "Multimodal Estimation of Distribution Algorithms,"
optimization algorithm (BACO) to address the balanced IEEE Trans. Cybern., vol. 47, pp. 636-650, 2017.
multiple traveling salesmen problem (BMTSP). Specifically, [14] L. Lu and T. Yue, "Mission-Oriented Ant-Team ACO for Min-Max
to simultaneously optimize the total path length of all MTSP," in Int. Conf. Inf., Commun. Eng., 2017, pp. 522-525.
[15] W. Liu, S. Li, F. Zhao, and A. Zheng, "An Ant Colony Optimization
salesmen and the longest path length to keep the path length Algorithm for the Multiple Traveling Salesmen Problem," in Proc.
balance, we have developed four ant selection strategies, IEEE Conf. Ind. Electron. Appl., 2009, pp. 1533-1537.
namely, Random Selection (RS), Shortest Biased Selection [16] M. Wang et al., "Ant Colony Optimization with an Improved
(SBS), Future Balance Biased Selection (FBBS) and Future Pheromone Model for Solving MTSP with Capacity and Time Window
Constraint," IEEE Access, vol. 8, pp. 106872-106879, 2020.
Shortest Biased Selection (FSBS). [17] Y. Harrath, A. F. Salman, A. Alqaddoumi, H. Hasan, and A. Radhi, "A
Experiments conducted on four popular TSP sets with Novel Hybrid Approach for Solving the Multiple Traveling Salesmen
different numbers of salesmen have demonstrated that the Problem," Arab J. Basic Appl. Sci., vol. 26, pp. 103-112, 2019.
proposed BACOs with the four selection strategies achieve [18] Q. Xu, Z. Mao, and Z. Jin, "Simulated Annealing-Based Ant Colony
Algorithm for Tugboat Scheduling Optimization," Math. Probl. Eng.,
much better performance than a state-of-the-art GA. In vol. 2012, pp. 1457-1461, 2012.
particular, the experimental results have shown that FSBS is [19] V. I. Lupoaie, I. A. Chili, M. E. Breaban, and M. Raschip, "SOM-
the most effective selection strategy, which helps ACO Guided Evolutionary Search for Solving MinMax Multiple-TSP," in
achieve the best performance in solving BMTSP among the Proc. IEEE Congr. Evol. Comput., 2019, pp. 73-80.
[20] H. Xu and C. Zhang, "The Research about Balanced Route MTSP
four selection strategies. Based on Hybrid Algorithm," in Int. Conf. Commun. Software Netw.,
2009, pp. 533-536.
REFERENCES [21] S. Yuan et al., "A New Crossover Approach for Solving the Multiple
[1] R. M. F. Alves et al., "Using Genetic Algorithms to Minimize the Travelling Salesmen Problem Using Genetic Algorithms," Eur. J. Oper.
Distance and Balance the Routes for the Multiple Traveling Salesman Res., vol. 228, pp. 72-82, 2013.
Problem," in Proc. IEEE Congr. Evol. Comput., 2015, pp. 3171-3178.

481

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PIAUI. Downloaded on December 30,2024 at 14:25:06 UTC from IEEE Xplore. Restrictions apply.