Received October 13, 2019, accepted October 23, 2019, date of publication October 29, 2019, date of current

version November 12, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2950214

Multi-Population Ant Colony Optimization

Algorithm Based on Congestion Factor
and Co-Evolution Mechanism
HAINAN ZHANG AND XIAOMING YOU
School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China
Corresponding author: Xiaoming You ([email protected])
This work was supported in part by the National Natural Science Foundation of China under Grant 61673258 and Grant 61075115, and in
part by the Shanghai Natural Science Foundation under Grant 19ZR1421600.

ABSTRACT To solve large-scale traveling salesmen problem (TSP) with better performance, we propose a
multi-population ant colony optimization algorithm based on congestion factor and co-evolution mechanism
(CCMACO). First, the congestion factor is introduced to control the number of ants on the path and
can improve the adaptability of CCMACO. Then, the sub-populations restructuring strategy is proposed
to balance the convergence speed and the diversity of solutions. Besides, the inter-specific competition
mechanism can be used to strengthen the optimal solution and to accelerate convergence speed. Finally,
the co-evolutionary strategy is used to interchange information among different sub-populations so as
to maintain the diversity of populations. For the purpose of verifying the optimization performance of
the CCMACO algorithm, CCMACO is compared with several improved multi-population ant colony
optimization algorithms in TSP. The experiment results show that the proposed CCMACO algorithm can
effectively obtain the best optimization value in solving TSP and it achieves better optimization ability and
stability.

INDEX TERMS Congestion factor, sub-populations restructuring strategy, inter-specific competition mech-
anism, co-evolutionary strategy, TSP.

I. INTRODUCTION the probability that ants will fall into a local optimum.
Traveling Salesman Problem (TSP) is a typical NP-Hard To limit too many pheromones, T.Stutzle et al. proposed
problem in mathematics, meaning that a businessman finds MAX-MIN Ant System (MMAS) [15]. As ant colony algo-
the shortest path to travel to all cities. Many optimization rithms evolve from single populations to multiple popula-
algorithms can solve TSP problems, such as Genetic Algo- tions, Twomey et al. [16] proposed multiple populations ant
rithm (GA) [1], Ant Colony Optimization Algorithm (ACO) colony algorithm.
[2], [3] Particle Swarm Optimization Algorithm (PSO) [4], Scholars proposed various improvement strategies based
Artificial Fish Swarm Algorithm (AFSA) [5], Artificial Bee on the ant colony algorithm. Xu et al. [17] proposed a new
Colony Algorithm (ABC) [6], [7] et al. ACO is a heuristic heuristic communication heterogeneous dual population ant
algorithm proposed by Dorigo et al. [8], which has paral- colony optimization for TSPs. Zheng [18] proposed an auto-
lelism and is easy to combine with other optimization meth- matic test case generation method of parallel multi-group
ods. The practical application has extended from TSP to adaptive ant colony algorithm. Zhao et al. [19] proposed a
all aspects of combinatorial optimization. Including robot new fault diagnosis method based on the principal component
path planning problems [9], [10] network routing problem analysis and the broad learning system for rotor system.
[11], [12] and secondary allocation problem [13] et al. Lu and Yue [20] proposed assigned ants to teams with
In the 90s, M.Dorigo proposed Ant Colony System Algo- mission-oriented approaches to enhance ant colony optimiza-
rithm (ACS) [14], adding a local search strategy to reduce tion algorithms. Deng et al. [21] proposed an improved
ant colony optimization algorithm based on the multi-
The associate editor coordinating the review of this manuscript and population strategy, co-evolution mechanism, pheromone
approving it for publication was Ran Cheng . updating strategy, and pheromone diffusion mechanism to

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
158160 VOLUME 7, 2019
H. Zhang, X. You: Multi-Population ACO Algorithm Based on Congestion Factor and Co-Evolution Mechanism

balance the convergence speed and solution diversity, and 3. The co-evolutionary mechanism exchanges pheromones
improve the optimization performance in solving the large- of dual populations in order to maintain the diversity of
scale optimization problem. Yu et al. [22] proposed a two- populations.
layer algorithm, namely ACO-A∗, by combining the ant
colony optimization with the A∗ search. Kurdi [23] proposed II. RELATED WORK
an ACS with a novel Non-Daemon Actions procedure for A. ANT COLONY OPTIMIZATION ALGORITHM
multiprocessor task scheduling in multistage hybrid flow 1) ACS ALGORITHM
shop. Zhou et al. [24] proposed an algorithm of a multi- There is a total of n cities in the proposed TSP. At the initial
objective multi-population ant colony optimization for a con- moment of the algorithm, m ants are randomly placed in n
tinuous domain. Zhao et al. proposed a new feature extraction cities. At the same time, the first element of each ant’s taboo
method based on EEMD and multi-scale fuzzy entropy for table is set to its current city. At this time, the amount of
motor bearing. Wu et al. [25] proposed an improved adap- pheromone on each path is equal and c is a smaller constant.
tive particle swarm optimization algorithm based on making Each ant independently selects the next city based on the
full use of the advantages of Alpha-stable distribution for amount of pheromone remaining in the path and the heuristic
solving multi-objective gate assignments. Ismkhan [26] pro- information (the distance between the two cities). At time t,
posed new strategies including effective representation and the probability pkij (t) of the ant k moving from the city i to the
heuristics, which speed up ACO and enable it to be applied city j is:
to large-scale instances. Deng et al. [27] proposed a genetic  α β
and ant colony adaptive collaborative optimization algorithm  P[τij (t)] [ηij (t)]

, j ∈ Jk (i)
[τis (t)]α [ηis (t)]β

for solving complex optimization problems. Mahi et al. [28] pkij (t) = (1)
proposed a new hybrid method is proposed to optimize  s∈Jk (i)

0 otherwise

parameters that affect the performance of the ACO algorithm
using Particle Swarm Optimization. Wu et al. [29] proposed where [Jk (i) = {1, 2, · · · , n} − tabuk ] indicates the set of
a new airport gate assignment method to effectively improve cities that ant k next allows selecting. The tabu list records
the comprehensive operation capacity and efficiency of hub the city where the ant k is currently passing. When all n cities
airport. Labrador Rivas et al. [30] proposed An extended are added to the tabuk , ant k completes a journey, and the
continuous domain ant colony optimization algorithm and a path that the ant k has passed is a feasible solution for the
hybrid algorithm (traditional ACO plus linear programming TSP. In formula (1), ηij (t) is a heuristic factor. α and β are
solver) to solve a mixed non-linear programming problem the control parameters.
with a high-level mathematical complexity. Wang et al. [31] When all the ants complete a tour, the pheromone on each
proposed an improved ACO algorithm approach for WSNs path is updated according to formula (2) and formula (3):
that use mobile sinks by considering CH distances. Orhan
and Güçlü [32] proposed an effective new hybrid ant colony τij = (1 − ρ)τij + 1τij (2)
algorithm based on crossover and mutation mechanism for Xm
no-wait flow shop scheduling with the criterion to minimize 1τij = 1τijk (3)
the maximum completion time. Uthayakumar et al. [33] k=1
proposed an ant colony optimization based financial crisis In here, ρ(0 < ρ < 1) indicates the evaporation coefficient
prediction model. of the pheromone on the path, 1 − ρ indicates the residual
As can be known from the above analysis of these related coefficient of the pheromone, 1τij represents the increment
works, the researchers have proposed various ACO algo- of the pheromone on the edge ij (city i to city j) after the end
rithms. These improved ACO algorithms are used to solve of this iteration.1τijk represents the amount of pheromone left
complex optimization problems and obtain better optimiza- by the k ant on the edge ij in this iteration. If ant k does not
tion results. But they still exist the slow convergence speed pass edge ij, the value of 1τijk is 0. 1τijk is updated by the
and are easy to fall into local optimal value. following formula:
In this paper, we focus on improving the accuracy of the 
solution on large-scale TSPs and reduce the running time  Q if the k ant used edge(i,j) in its tour
of the algorithm. The main contributions of this paper are 1τijk = Lk (4)
0 otherwise

summarized as follows:
1. The congestion factor is proposed to control the number where Q is the information strength and Lk is the path length
of ants on the path and the congestion threshold was adjusted of the k-th ant in the current loop.
through the change of the iteration cycle.
2. The pheromone incentive or penalty mechanism is intro- 2) MMAS ALGORITHM
duced can reward the population with low fitness and punish In order to avoid the algorithm convergence too fast and stag-
the population with high fitness to accelerate the convergence nate, the MMAS algorithm limits the pheromone to a certain
speed of the algorithm. range: [τmin , τmax ]. If all the ants have completed a round trip,

VOLUME 7, 2019 158161

 H. Zhang, X. You: Multi-Population ACO Algorithm Based on Congestion Factor and Co-Evolution Mechanism

the pheromone on each path of the optimal solution is updated ants to explore repeat paths rather than exploring other paths,
according to formula (5), so as to accelerate the convergence. which can improve the global searching capacity. As the num-
ber of iterations increases, the congestion threshold increases
τi,j (t + n) = (1 − ρ) × τi,j (t) + 1τi,j
best
ρ ∈ (0, 1) (5)
and more ants are scheduled to select the optimal path so as
1τi,j
best
= 1/Lbest (6) to accelerate algorithm convergence.
where Lbest is the globally best path and the Lbest in the The calculation formula for the congestion degree of the
following improved algorithm uses the optimal path in this path is described as follows:
round of loop. τij (t)
qij = P (8)
3) TSP τij (t)
The solution to the TSP problem is to traverse the path i6=j
length of l cities, so the fitness function is updated by the wherePτij (t) is the pheromone amount on the ant selection
following formula: path, τij (t) is the total pheromone amounts on all paths.
l−1 i6=j
The congestion factor is described as follows:
X
f = D (xs , xs+1 ) + D (xl , x1 ) (7)
−t
s=1
σ (t) = 1 − e s(s−1) (9)
where D (xs , xs+1 ) is the distance between the s visited city
and the s + 1 visited city. The shorter the path, the smaller the where s is the number of cities and e is naturally constant.
fitness. x1 , xl are the first city and the last city. When the ant chooses the path to go next, the congestion of
the path will be computed. If qij < σ (t), the ant chooses this
III. MULTI-POPULATION ANT COLONY OPTIMIZATION path, otherwise, reselects the path that can be taken.
ALGORITHM BASED ON CONGESTION FACTOR
AND CO-EVOLUTION MECHANISM 3) SUB-POPULATIONS RESTRUCTURING STRATEGY
A. INTRA-SPECIES COMPETITION AND When all sub-populations complete a path, the sub-group
COOPERATION MECHANISM reconstruction strategy is implemented and the last two sub-
1) ADAPTIVE EVOLUTION STRATEGY population are disrupted restructuring and keeping the num-
From the analysis of section II, so we can draw such a ber of n sub-populations within the populations unchanged.
conclusion that ACS and MMAS have good complementarity In the first round of iterative evolution, the restructuring sub-
and can achieve a good balance between convergence speed populations integrate the optimal path of the existing elite
and diversity. In the process of evolution, natural biological sub-populations and selects the corresponding nodes.
populations follow the rules of survival of the fittest. Based The location of the restructuring sub-population is
on this principle, we set the population A and population B described as follows:
which adopt the ACS and MMAS algorithms respectively, θj = θbest + r(θbest − θ1 ) (10)
and each of them optimize in the early iteration. After reach-
ing a certain number of algebras, the two-populations coop- where θbest and θ1 are the location of the current optimal ant
eration and co-evolution accelerate the convergence speed of and the location of the randomly selected ant respectively. r is
the algorithm. the exploration radius. To avoid prematurely, the restructur-
In order to evolve the populations toward superiority, ing sub-population explores the space around the elite sub-
which are divided into n sub-populations with the same num- population in the early stages of evolution. At the same time,
ber of ants. Each sub-population has the same state, adopts an a protection strategy is applied, including the principle of not
adaptive evolution strategy, and survives the inferior, which participating in the first competition. The strategy ensures
promotes the population. The n sub-populations are placed at that the restructuring sub-populations explore the solution
different starting points for searching. After certain iterations, space adequately in a good evolutionary environment.
the inferior sub-populations are replaced by the dominant
sub-populations. Then all populations will be reorganized, B. INTER-SPECIFIC COMPETITION MECHANISM
which is followed by the next iteration. When the sub-populations evolve a certain number of
iterations alone, the dual populations adopt a competitive
2) DYNAMIC SCHEDULING STRATEGY mechanism, giving appropriate pheromone incentives or
In exploring feasible paths, the sub-populations tend to con- penalties to both the winner and the loser, which strengthens
centrate on searching for the same area, causing waste of the competitive population and speeds up the convergence of
feasible domain resources and increasing the probability of the algorithm.
falling into a local optimum. To address it, the congestion When the competition period satisfies Tw%D = 0, the dual
factor of the artificial fish swarm algorithm is introduced populations are ranked by competition. Where w is the total
and the number of ants passing through the controlling path number of sub-populations retained by the dual population
adjusted adaptively according to the iterations. In the early in the last inter-population competition, and D is a constant
stage, we set a small congestion threshold, avoiding excessive parameter equaling to 5. Comparing the optimal fitness of

158162 VOLUME 7, 2019

H. Zhang, X. You: Multi-Population ACO Algorithm Based on Congestion Factor and Co-Evolution Mechanism

population A and population B, the evaluation factor E is their pheromone matrix according to formula (14).
set to evaluate the competitiveness of populations A and B,
Pheromone(a,i)  Pheromone(b,i) (14)
which are described as follows:
|f1best − f2best | where (a, i) is the i sub-population of population A and (b, i)
E= (11) is the i sub-population of population B. Because each sub-
n · max fmax − fav , fav − fopt

population has different current optimal paths before com-
where, f1best , f2best are the optimal fitness of dual populations munication, the pheromone on the path after communication
respectively, fmax , fav , fopt represent the maximum, average is easy to exceed the limit. Therefore, the pheromone matrix
and theoretical optimal values of the fitness of the dual is used to adjust the pheromone matrix after communication
populations. within the limits.
After evaluating the competitiveness of the two, each pair
is given a certain pheromone incentive, and the incentive D. THE MODEL AND STEPS OF THE CCMACO ALGORITHM
function is described as follows: The model of the CCMACO algorithm based on intra-
( species competition and cooperation mechanism, dynamic
3Pi E ∈ [0, 1%]
Pi = (12) scheduling strategy interspecific competition mechanism,
7Pi E ∈ (1%, 5%] co-evolutionary strategy is shown in Figure 2.
m
P
where, the initial value Pi = 2m/ fk , fk is the fitness of
k=1
k ant. In the next round of iterative evolution of the dual
population, a global positive pheromone incentive is given
to the competition winner, conversely, a pheromone nega-
tive incentive is given. The pheromone update formula is
described as follows:
τij = (1 − ρ) · τij + ρ · 4τij ± Pi (13)

C. CO-EVOLUTIONARY STRATEGY
In order to achieve efficient cooperation between the two
populations and maximize the benefits, the dual populations
adopt a coordinated evolution model. It can be seen from
the previous section that the average fitness of each sub-
population is fvi , and each sub-population is ranked according
to the fvi value from high to low. The sub-population with
the highest fitness cooperates with the sub-population with
the lowest fitness in another population and communicates in
turn. The evolution patterns of population A and B are shown FIGURE 2. The model of the CCMACO algorithm.
in Figure 1.
The steps of the CCMACO algorithm are described as
follows.
Step 1. Divide the population into n sub-population and
place them at different starting points.
Step 2. Initialize the parameters of the CCMACO algo-
rithm. These parameters include the number of ants (m),
the maximum number of iterations (T ), the parameter (α, β),
FIGURE 1. Populations co-evolution model. volatility coefficient (ρ), etc.
Step 3. Each sub-population independently execute the
The co-evolution strategy is a new kind of evolutionary search process. and the next node is selected according to
strategy based on co-evolution theory in recent years. The formula (1).
sub-populations with high fitness are based on the path infor- Step 4. Calculate the congestion degree on each path. Elim-
mation provided by the sub-population with low fitness, open inate the sub-population according to formula (8) and (9).
up new solution space exploration. The sub-populations with Step 5. Randomly reorganize into a new sub-population
low fitness fit into the pheromone on the sub-population path but the first iteration search according to formula (10).
with high fitness, preventing the ants from falling into local Step 6. Locally update the pheromone according to for-
optimum during the search process. mula (4).
The paired sub-population of population A and population Step 7. Globally update pheromone according to formula
B are updated with the element τij of row i and column j of (2) and (5).

VOLUME 7, 2019 158163

 H. Zhang, X. You: Multi-Population ACO Algorithm Based on Congestion Factor and Co-Evolution Mechanism

Step 8. Evaluate the competitiveness of populations to describe and compare the experimental results. The Opt
according to formula (11) and motivate or penalize popula- represents the theoretical optimal value of the TSP. The best
tions according to formula (12) and (13). value represents the obtained minimum value in the simu-
Step 9. Populations coevolution, exchange pheromone lation test of 15 times and the average value represents the
according to formula (14). average value of 15 times. The experiment results are shown
Step 10. Determine whether the maximum number of iter- in Table 2.
ations (T ) is achieved or the obtained solution has met the The experiment results of the basic ACS algorithm,
requirement. If this end condition does not meet, then execute the MMAS algorithm and the proposed CCMACO algorithm
Step 3 so as to start a new evolution. in solving TSP standard instances are shown in Table 2. For
three optimization algorithms, the experiment results show
IV. EXPERIMENT AND SIMULATION that the CCMACO algorithm can obtain the best optimization
A. EXPERIMENT ENVIRONMENT AND PARAMETERS values in solving TSP standard examples within 100. It shows
In order to demonstrate the optimization performance of that the CCMACO algorithm has better optimization ability
the proposed CCMACO algorithm, twelve TSP standard than the basic ACS algorithm and the MMAS algorithm
instances from TSPLIB standard library are selected in this in solving these TSP standard instances. From the average
paper. According to the characteristic of TSPLIB, the dis- value, the CCMPACO algorithm can also obtain the best
tance between any two cities is computed by the Euclidian average value. It shows that the optimization performance
distance and then take the integer. The basic ACS algorithm of the CCMACO algorithm is a more obvious advantage.
and MMAS algorithm are selected to compare the optimiza- It can be seen that the standard deviation of the proposed
tion performance with the proposed CCMACO algorithm. CCMACO algorithm in solving the TSP is the smallest value
The experiment environments are Matlab2016b, the Pentium for eil51, berlin52, eil101, ch150, rat195, which shows that
CPU i5, 8.0GB RAM with Windows10. The values of param- the CCMACO algorithm has better stability than ACS algo-
eters in these algorithms could be a complicated problem rithm and MMAS algorithm.
because change parameters could affect the optimum value.
In the simulation experiments, the alternative values were
tested and modified for some functions to obtain the most
reasonable initial values of these parameters. These selected
values of the parameters take on the optimal solution and
the most reasonable running time of these algorithms to effi-
ciently complete the problem-solving. So the most reasonable
initial values of these parameters are obtained by testing and
modifying. The obtained initial values of these parameters are
shown in Table 1.

TABLE 1. The set of parameters.

FIGURE 3. Best paths for small scale TSP instances found by CCMACO.

Figure 3 illustrates the CCMACO algorithm found the best

routes for TSP and their costs (their route lengths).
In order to further demonstrate the optimization perfor-
mance of the proposed CCMACO algorithm, this section
analyzes the convergence speed of the algorithm, selects two
representative cities, and compares the number of iterations
required for the three algorithms to converge to the shortest
B. EXPERIMENT RESULTS AND ANALYSIS path. As can be seen from Figure 4 that the CCMACO
In the simulation experiment, for each TSP standard instance, algorithm has the fastest convergence rate and the shortest
the basic ACS algorithm, MMAS algorithm, and the proposed search path, indicating that the CCMACO algorithm has good
CCMACO algorithm are all executed 15 times. The indexes performance in global optimization.
of best value, average value (Avg), relative error value (Re) Ten typical cities are selected to visualize the iterative data,
and standard deviation (Std) value of the 15 results are used as shown in Figure 5. In each TSP problem, the number of

158164 VOLUME 7, 2019

H. Zhang, X. You: Multi-Population ACO Algorithm Based on Congestion Factor and Co-Evolution Mechanism

TABLE 2. Comparison of different algorithms.

VOLUME 7, 2019 158165

 H. Zhang, X. You: Multi-Population ACO Algorithm Based on Congestion Factor and Co-Evolution Mechanism

TABLE 3. Comparison of different algorithms.

FIGURE 4. Convergence rate comparison.

FIGURE 6. Best paths for each TSP instance found by CCMACO.

C. COMPARISON WITH OTHER MULTI- COLONIES

ANT COLONY ALGORITHMS
For the purpose of demonstrating the optimization perfor-
mance of the proposed CCMACO algorithm, the ICMPACO
algorithm [21], PSO-ACO-3Opt algorithm [28] and HHACO
algorithm [17] are used to solve the TSP problem. These
FIGURE 5. Comparison of iterations. parameters are set the same in Table 1. The experiment results
are shown in Table 3.
iterations of ACS and MMAS are obviously more than the As showing in Table 3, it can clearly see the optimiza-
CCMACO algorithm. It shows that CCMACO algorithm has tion performance and effect of the basic ICMPACO algo-
a distinct advantage in the speed of optimization. rithm, PSO-ACO-3Opt algorithm, HHACO algorithm and the

158166 VOLUME 7, 2019

H. Zhang, X. You: Multi-Population ACO Algorithm Based on Congestion Factor and Co-Evolution Mechanism

TABLE 4. Comparison of different strategies by variable-controlling approach.

TABLE 5. Comparison of different algorithms.

TABLE 6. Statistical result of CCMACO in different instances.

CCMACO algorithm in solving TSP. For four optimization The experiments have been repeated 15 times in
algorithms, the experiment results show that the CCMACO kroB150 and rat195. The variable controlling approach can
algorithm can obtain the best optimization values and the be used to validate the performance of each strategy and its
minimum average values. It shows that the CCMACO algo- contribution to CCMACO. The statistic results of the three
rithm has an obvious positive effect on competition coopera- algorithms have been shown in Table 4. Both in kroB150 and
tion mode. rat195, algorithm_1, which is CCMACO without dynamic

VOLUME 7, 2019 158167

 H. Zhang, X. You: Multi-Population ACO Algorithm Based on Congestion Factor and Co-Evolution Mechanism

scheduling strategy, has a good performance. In algorithm_2, [2] H. Zhang, X. You, and S. Liu, ‘‘Ant colony optimization algorithm based on
there is no communication between population A and pop- dynamic scheduling strategy and competition mechanism,’’ Microelectron.
Comput., vol. 36, no. 7, pp. 36–42, 2019, doi: 10.19304/j.cnki.issn1000-
ulation B but the results are very accurate. CCMACO mix 7180.2019.07.008.
algorithm_1 with algorithm_2 and the statistic results indi- [3] H. Zhu, X. You, and S. Liu, ‘‘Multiple ant colony optimization based on
cate that the complete CCMACO has a better performance pearson correlation coefficient,’’ IEEE Access, vol. 7, pp. 61628–61638,
2019, doi: 10.1109/ACCESS.2019.2915673.
than others both in accuracy and stability. [4] B. Cheng, H. Lu, Y. Huang, and K. Xu, ‘‘An improved particle swarm
Figure 6 illustrates CCMACO algorithm found the best optimization algorithm based on cauchy operator and 3-opt for TSP,’’ in
routes for TSP and their costs (their route lengths). Proc. 17th Int. Conf. Parallel Distrib. Comput. Appl. Technol., Dec. 2016,
pp. 177–182.
In the cause of describing the optimization performance of [5] T. Fei, L. Zhang, Y. Li, Y. Yang, and F. Wang, ‘‘The artificial fish swarm
the CCMACO algorithm, the HHACO algorithm [17] and the algorithm to solve traveling salesman problem,’’ in Proc. Int. Conf. Com-
IPGA algorithm [1] is used as the comparison algorithm and put. Sci. Inf. Technol., 2014, pp. 679–685.
the experimental comparison results are shown in Table 5. [6] L. Li, Y. Cheng, L. Tan, and B. Niu, ‘‘A discrete artificial bee colony
algorithm for TSP problem,’’ in Proc. Int. Conf. Intell. Comput., 2011,
The IPGA combines PSO and GA and its performance is pp. 566–573.
worse than other algorithms. The sub-populations of HHACO [7] Y. Zhong, J. Lin, L. Wang, and H. Zhang, ‘‘Hybrid discrete artificial
are AS and IACS, which is the improvement of ACS. It has bee colony algorithm with threshold acceptance criterion for traveling
salesman problem,’’ Inf. Sci., vol. 421, pp. 70–84, Dec. 2017.
similar performance with CCMACO but cannot find the best [8] M. Dorigo, V. Maniezzo, and A. Colorni, ‘‘Ant system: Optimization by
solution of the three TSP instances. a colony of cooperating agents,’’ IEEE Trans. Syst., Man, Cybern. B,
As can be seen from the comprehensive result of TSP Cybern., vol. 26, no. 1, pp. 29–41, Feb. 1996, doi: 10.1109/3477.484436.
[9] X. Chen, P. Zhang, G. Du, and F. Li, ‘‘Ant colony optimization based
examples, the proposed CCMACO algorithm can keep rela- memetic algorithm to solve Bi-objective multiple traveling salesmen prob-
tive error below 0.5%, and obtain the ideal standard deviation lem for multi-robot systems,’’ IEEE Access, vol. 6, pp. 21745–21757,
results. We can know that the CCMACO algorithm takes on 2018.
[10] H. Yang, J. Qi, Y. Miao, H. Sun, and J. Li, ‘‘A new robot navigation algo-
better optimization performance and has the best precision in rithm based on a double-layer ant algorithm and trajectory optimization,’’
solving TSP examples. IEEE Trans. Ind. Electron., vol. 66, no. 11, pp. 8557–8566, Nov. 2019, doi:
Table 6 shows the statistical results of 10 different TSP 10.1109/TIE.2018.2886798.
[11] K. Mahmood, S. Kamal, and H. M. Khan, ‘‘Dynamic optimization of
instances, which proves CCMACO has better solution stabil-
network routing problem through ant colony optimization (ACO),’’ IEEE
ity and precision. Trans. Cybern., vol. 28, no. 5, pp. 629–640, Aug. 2012.
[12] A. Nayyar and R. Singh, ‘‘IEEMARP—A novel energy efficient multipath
routing protocol based on ant Colony optimization (ACO) for dynamic sen-
V. CONCLUSION sor networks,’’ Multimedia Tools Appl., May 2019, doi: 10.1007/s11042-
In this paper, a new multi-population ant colony optimization 019-7627-z.
(CCMACO) algorithm based on the congestion factor, sub- [13] S. Oliveira, M. S. Hussin, A. Roli, M. Dorigo, and T. Stützle, ‘‘Anal-
ysis of the population-based ant colony optimization algorithm for the
populations restructuring strategy, inter-specific competition TSP and the QAP,’’ in Proc. IEEE Congr. Evol. Comput., Jun. 2017,
mechanism and co- evolutionary strategy is proposed to solve pp. 1734–1741.
the large-scale complex optimization problems. CCMACO [14] M. Dorigo and L. M. Gambardella, ‘‘Ant colony system: A cooperative
algorithm is compared with the basic ACS algorithm, MMAS learning approach to the traveling salesman problem,’’ IEEE Trans. Evol.
Comput., vol. 1, no. 1, pp. 53–66, Apr. 1997.
algorithm, and other multi-population ant colony optimiza- [15] T. Stutzle and H. Hoos, ‘‘MAX-MIN ant system and local search for the
tion algorithms in solving the traveling salesmen problem. traveling salesman problem,’’ in Proc. IEEE Int. Conf. Evol. Comput.,
The proposed CCMACO algorithm can obtain the best opti- Apr. 1997, pp. 309–314.
[16] C. Twomey, T. Stützle, M. Dorigo, M. Manfrin, and M. Birattari, ‘‘An anal-
mization value in solving these TSP standard examples. ysis of communication policies for homogeneous multi-colony ACO algo-
Therefore, the proposed CCMACO algorithm takes on better rithms,’’ Inf. Sci., vol. 180, no. 12, pp. 2390–2404, Jun. 2010.
optimization ability and stability than the ACO algorithm and [17] M.-L. Xu, X.-M. You, and S. Liu, ‘‘A novel heuristic communication
heterogeneous dual population ant colony optimization algorithm,’’ IEEE
multi-population algorithm. Access, vol. 5, pp. 18506–18515, 2017.
Because the CCMACO algorithm exists the longer com- [18] T. Zheng, ‘‘Automatic test case generation method of parallel multi-
putation time in solving a complex optimization problem, population self-adaptive ant colony algorithm,’’ in Recent Developments
in Intelligent Computing, Communication and Devices, S. Patnaik and
the CCMACO algorithm needs to further be studied in order V. Jain, Eds. Singapore: Springer, 2018, pp. 469–476.
to reduce the time complexity. In future work, we will con- [19] H. Zhao, J. Zheng, J. Xu, and W. Deng, ‘‘Fault diagnosis method based on
tinue to study deeply on parameter optimization and solving principal component analysis and broad learning system,’’ IEEE Access,
vol. 7, pp. 99263–99272, 2019.
large-scale problems.
[20] L.-C. Lu and T.-W. Yue, ‘‘Mission-oriented ant-team ACO for min–max
MTSP,’’ Appl. Soft Comput., vol. 76, pp. 436–444, Mar. 2019.
ACKNOWLEDGMENT [21] W. Deng, J. Xu, and H. Zhao, ‘‘An improved ant colony optimization algo-
rithm based on hybrid strategies for scheduling problem,’’ IEEE Access,
The author would like to thank his senior Mingle Xu who had vol. 7, pp. 20281–20292, 2019.
given him some advice to make the article be much better. [22] X. Yu, W.-N. Chen, T. Gu, H. Yuan, H. Zhang, and J. Zhang, ‘‘ACO-A*:
Ant colony optimization plus A* for 3-D traveling in environments with
dense obstacles,’’ IEEE Trans. Evol. Comput., vol. 23, no. 4, pp. 617–631,
REFERENCES Aug. 2019.
[1] H. Zhou, M. Song, and W. Pedrycz, ‘‘A comparative study of improved [23] M. Kurdi, ‘‘Ant colony system with a novel Non-DaemonActions proce-
GA and PSO in solving multiple traveling salesmen problem,’’ Appl. Soft dure for multiprocessor task scheduling in multistage hybrid flow shop,’’
Comput., vol. 64, pp. 564–580, Mar. 2018. Swarm Evol. Comput., vol. 44, pp. 987–1002, Feb. 2019.

158168 VOLUME 7, 2019

H. Zhang, X. You: Multi-Population ACO Algorithm Based on Congestion Factor and Co-Evolution Mechanism

[24] J. Zhou, C. Wang, Y. Li, P. Wang, C. Li, P. Lu, and L. Mo, ‘‘A multi- HAINAN ZHANG was born in 1993. He is cur-
objective multi-population ant colony optimization for economic emission rently pursuing the M.S. degree with the Shanghai
dispatch considering power system security,’’ Appl. Math. Model., vol. 45, University of Engineering Science. His research
pp. 684–704, May 2017. interests include intelligent algorithms, path plan-
[25] D. Wu, H. Zhao, X. Yang, J. Xiong, M. Sun, and B. Li, ‘‘Study on an ning of mobile robots, and embedded systems.
improved adaptive PSO algorithm for solving multi-objective gate assign-
ment,’’ Appl. Soft Comput., vol. 59, pp. 288–302, Oct. 2017.
[26] H. Ismkhan, ‘‘Effective heuristics for ant colony optimization to han-
dle large-scale problems,’’ Swarm Evol. Comput., vol. 32, pp. 140–149,
Feb. 2017.
[27] W. Deng, H. Zhao, L. Zou, G. Li, X. Yang, and D. Wu, ‘‘A novel col-
laborative optimization algorithm in solving complex optimization prob-
lems,’’ Soft Comput., vol. 21, no. 15, pp. 4387–4398, Aug. 2017, doi:
10.1007/s00500-016-2071-8.
[28] M. Mahi, Ö. K. Baykan, and H. Kodaz, ‘‘A new hybrid method based on
particle swarm optimization, ant colony optimization and 3-opt algorithms
for traveling salesman problem,’’ Appl. Soft Comput., vol. 30, pp. 484–490,
May 2015.
[29] D. Wu, S. Meng, H. Zhao, L. Bo, and C. Wang, ‘‘Study on an airport
gate assignment method based on improved ACO algorithm,’’ Kybernetes,
vol. 47, no. 1, pp. 20–43, 2018.
[30] A. E. L. Rivas, L. A. G. Pareja, and T. Abrāo, ‘‘Coordination of distance
and directional overcurrent relays using an extended continuous domain
ACO algorithm and an hybrid ACO algorithm,’’ Electr. Power Syst. Res., XIAOMING YOU received the M.S. degree in
vol. 170, pp. 259–272, May 2019.
computer science from Wuhan University, in 1993,
[31] J. Wang, J. Cao, R. Sherratt, and J. H. Park, ‘‘An improved ant colony
and the Ph.D. degree in computer science from
optimization-based approach with mobile sink for wireless sensor net-
works,’’ J. Supercomput., vol. 74, no. 12, pp. 6633–6645, Dec. 2018. the East China University of Science and Tech-
[32] O. Engin and A. Güçlü, ‘‘A new hybrid ant colony optimization algorithm nology, in 2007. She is currently a Professor with
for solving the no-wait flow shop scheduling problems,’’ Appl. Soft Com- the Shanghai University of Engineering Science.
put., vol. 72, pp. 166–176, Nov. 2018. Her research interests include swarm intelligent
[33] U. J, N. Metawa, K. Shankar, and S. K. Lakshmanaprabu, ‘‘Financial crisis systems, distributed parallel processing, and evo-
prediction model using ant colony optimization,’’ Int. J. Inf. Manage., to lutionary computing.
be published.

VOLUME 7, 2019 158169