Você está na página 1de 14

Jennifer Marx

May 01, 2012


Spring 2012

Min/Max Multiple Traveling Salesmen Problem: Modified GA with a 2-Opt Mutation

1. Introduction
Although there exists a great deal of research dealing with the Traveling Salesmen Problem
(TSP), only limited attention has been given to the Multiple Traveling Salesmen Problem
(MTSP) [1]. Traditionally, the MTSP is defined as m-salesmen visiting n-cities exactly once
while minimizing the overall total distance traveled. The min/max variation uses a different
objective (cost) function of minimizing the distance of the salesman with the longest path.
Min/max MTSP has many useful applications because it helps balance the work-load among the
salesmen, which leads to similar mission completion times. Parlaktuna [4] explains how Vehicle
Routing Problems utilize the MTSP in path planning. The planning of autonomous homogenous
robots deployment demonstrates one of the general applications for the MTSP [1]. Real-world
examples may include deployment of multiple sinks in a wireless sensor network, land-mine
exploration [4], or transporting loads from one node or another [4].

Heuristic approaches remain the most popular solution methods for MTSP, because they
scale better than the exact solution methods [1]. Genetic algorithms (GAs) are a type of
evolutionary based algorithm, which are stochastic and model natural phenomena of genetic
inheritance and Darwinian survival of the fittest [5]. Research shows GAs effectiveness at
finding a near optimal solution to the min/max MTSP problem [2]. Even though GAs possess
strong global searching ability, they can have slow convergence or fall into a local minimum
[10]. GAs using mutation, crossovers, and a form of elitism perform better than GAs that do not
1

utilize all three [3]. While there has been previous research on improving the GA with a local
search [10,11,12], the research does not utilize the most effective GA algorithms available. To
the authors knowledge, there has not been any research towards solving the min/max MTSP
problem with a GA utilizing local search methods.

The paper is organized with Section 2 providing a background into GAs and providing
explanations for all operators used in the algorithm. Section 3 explains the modified two-part
chromosome algorithm. Section 4 contains a comparison of three different GAs though a
Matlab simulation, showing the improvements the modified GA algorithm creates. The paper
concludes with section 5, with possibilities of further studies.

2. GAs Background
GAs contain five components: genetic representations called chromosomes, ways to
create an initial population of potential solutions, a fitness function, operators to create offspring,
and values of needed parameters such as population size, mutation rates, etc. [5]

2.1 Chromosome representation


Chromosome representation plays a large role in the effectiveness and efficiency of
finding a near-optimal solution. There exist three different types of chromosome representations,
which have been studied and compared [6]. Previous research proves the two-part chromosome
method as the most efficient and effective method [6], therefore the author utilizes this approach.

The figures 1, 2, and 3 by Carter and Ragsdale [6] demonstrate the differences between
the three types. [8] The one chromosome technique [8] in Figure 1 partitions the salesmans
routes by m-1 negative integers, where m is the number of salesmen. For example, cities 2, 5, 14,

6 belong to salesmen 1. There are (n + m 1)! possible solutions to the problem with many
possible redundant options.

The two chromosome technique [9] in Fig. 2 shows an additional chromosome which
keeps track of the salesman assigned to each city. For examples, salesman 2 follows a route of 2,
8, 12, and 9. This technique has

possible solutions because of possible redundant

solutions [6].

Lastly, the two-part chromosome in Figure 3 maintains the number of cities per salesman
in the second part of the chromosome. For example, the first 4 cities, 2 , 5, 14, and 6 belong to
Salesman 1. It has been shown the solution space of the two-part chromosome decreases to
. The solution space size contributes the most to the efficiency of the two-part
chromosome method.

2.2 Initializing Population


Even though initial population can be seeded, the typical approach is a random generation
[5].
In order to maintain genetic diversity, choosing a large enough pool is very important. If the
initial population pool is too small, early convergence may be an issue. Each salesman receives a
random route.

2.3 Fitness Function


The MTSP min/max variation dictates the fitness function. The fitness function evaluates
each chromosome in the randomly selected population, finding the salesman with the longest
route. Comparing the distances of the maximum routes, the shortest distances are given a
preference to continue onwards. The most popular methods are utilizing either a Roulette Wheel
or Tournament Style selection method. [5] The Roulette Wheel assigns a percentage to the
evaluated populations, with the best population having a higher percentage of selection. The
Tournament Style simply choses the most fit chromosome or chromosomes to create offspring.

2.4 Population Generators (Crossovers, Mutations, and 2-Opt)


From the chosen chromosomes (parents), new offspring are created. The two main
approaches for creating new populations are utilizing crossovers and/or mutations. A certain
probability is assigned to each change, so crossovers or mutations may not occur to all selected
individuals. [5]

2.4.1 Crossovers Operators


Crossover operators focus on similarities between chromosomes [5], which keep the best
chromosomes in the population. Many different crossover operators exist [5], which change the
4

routing portions of the chromosomes. For the current work, the two-point central crossover and
cycling crossover are utilized because they showed the most promise in new research [7]. In
Figure 4, [7], two random numbers select cutting positions Parent A. The elements between cut 1
and cut 2 reorder according to their position in Parent B. Offspring B works the same, where it
copies Parent B, except the genes between cut 1 and cut 2 reorder according to their position in
parent A.

Fig. 4 Two-Point Crossover Operator [7]

Figure 5 demonstrates the cycling crossover method. [7] A random position chosen in
parent A swaps with the same position in parent B. Then the element in parent B is located in
parent A to find to new swapping positions. The swapping and finding continue until the starting
element from parent A is located in parent B.

Fig. 5 Cycling Crossover Operator [7]

In the two-part chromosomes, the second part of the chromosomes containing the
assigned cities may go through crossovers and mutations also. The current work uses two
different arithmetic crossovers from Chens [7] promising research. Equations 1 and 2 take two
random chromosomes and find the weighted-sum values at position k in both parents. Alpha is a
random value between (0, 1). Equation 3 and 4 is another arithmetic crossover [7]. The two
chromosomes mate at position k by combining, scaling, and then adding. The algorithm checks
the chromosomes for feasibility by making sure the summation of each city routes equals n
cities. If not, cities are randomly distributed or removed between the salesmen until n-cities are
visited for both chromosomes.

2.4.2 Mutations
Mutations allow for genetic diversity, which aids in preventing early convergence [5].
Similar to the crossovers, there are different options for mutation operators. Two simple
6

mutations are the inverse and swap, which changes the 1st part of the chromosome in the current
work. Figure 6 illustrates the inverse operator [7]. The genes within two random points reverse
their order. The swapping mutation simply chooses two random positions within a chromosome
and switches the elements.

Fig. 6 Inverse Mutation [7]

In the two-part chromosome, the second part of the chromosome may mutate. Cities from
one randomly selected salesman are moved to another random salesman, Formula 7 and 8 [7].
Similar to the crossover procedures, the second part of the chromosome checks for errors. If
solutions are not feasible, then the algorithm adds or reallocates cities randomly to salesmen.

2.4.3 Two-Opt Operator


The 2-opt algorithm is a local search algorithm. The operator minimizes each salesmans route
by deleting two edges and reconnecting those cities by adding two new edges. The 2-opt
algorithm rids a route of self-crossing portions [11]. See Figure 7 for illustration.

Fig. 7 Two-Opt Algorithm [11]

2.5 Parameter in Algorithm


GAs allow for many different variations, which will affect the performance of the
algorithm [5]. Chosen parameters include population pool size, selection size, probabilities for
crossovers, mutations, chromosome selections, and other fitness function requirements. The
fitness function requirements may include a min or max number of cities a salesman visits or
timing constraints. The convergence parameter dictates the number of cycles the GA will process
a best fit solution to end the algorithm. If the convergence number is too low, the GA will
converge to a local minimum. If the convergence number is too high, the GA will process for an
unnecessary amount of time. When choosing parameters, there does not seem to be a best
method in choosing the strongest parameters besides trial and error [7]. In the current work, the
algorithm uses trial and error methods with previous research in choosing parameters [7].

3. Modified Two-Part Chromosome Algorithm

Step 1: Create a random population pool, p. Select x random number of chromosomes to


evaluate. Choose the convergence parameter and global minimum.

Step 2: Using the fitness function, evaluate the selected chromosomes. From the Tournament
approach, select the top two most-fit chromosomes to continue as parents. So there will exist a
8

number one chromosome and a number two chromosome. Update global minimum with the
maximum distance a salesmen travels from number one chromosome.

Step 3: Generate new population group from operators. Choose a probability for each possible
change, two central point crossover, cycling crossover, inverse mutation, and swapping mutation
to occur on the best and second best solution. The 2-opt algorithm iterates on each best solution,
until it becomes the global minimum. When the best solution becomes the global minimum, the
2-opt algorithm occurs on the second best solution with some defined probability.

Step 4: Repeat step 2. If the best solutions fitness exceeds the global minimum, repeat step 3
with new best solution. Otherwise, retain previous best solution, and repeat step 3.

Step 5: If the global minimum remains the same for convergence parameter the algorithm stops
with the best solution to the min/max MTSP.

4. Experiment
4.1 Set-Up
An average was taken from 30 simulations, which ran in Matlab 7 on an Intel Core
i5 CPU, 3.33 GHz. The simulations compared single, 2-part, and modified 2-part chromosome
methods. Two chromosome method was left out because there have been studies showing it
under performing the single and 2-part methods [6]. Cities sizes had sparse density at 25 cities,
low medium density at 50 cities, medium density at 100 cities, and high density at 150 cities.
Initial population sizes were approximately double the city size in order to ensure a diverse
population. Each salesman has to visit at least two cities. Matlab produced the city maps with
the rand function. Crossover rate was set at 0.9, and mutation rates set at 0.5. The 2-opt mutation
rate for the second best solution is set at 0.5. These rates were derived from test simulations
9

yielding the best results. Convergence parameter was set at 1,000 cycles. Iterations were the
number of new best solutions, which were processed out. The higher the number of iterations,
the longer the GAs ran.

4.2 Results
Figures below provide the results from each set of simulations. The modified two-part
chromosome algorithm overall performed well in the various density scenarios. Performance
exceled most when the salesmen were fewer, ie m=2 or m=5. Both the maximum distance and
iterations were significantly less than the one and two-part chromosome methods. Also, the
modified algorithm performed best in the sparse density scenario. When the number of salesmen
were higher, the improvements in final distances were less drastic. However, the modified
algorithm always produced a result in less number of iterations.

25 Cities
One Chromosome
Iterations
Two-Part Chromosome
Iterations
Modified Two-Part Chromosome
Iterations

m=2
2150
1990
2142.5
1990
2036
1439

m=5
1655.3
2030
1574.9
1936
1554.4
1897

m=10
1333.2
1987
1319.9
2379
1307
2179

10

50 Cities
One Chromosome
Iterations
Two-Part Chromosome
Iterations
Modified Two-Part Chromosome
Iterations

100 Cities
One Chromosome
Iterations
Two-Part Chromosome
Iterations
Modified Two-Part Chromosome
Iterations

m=2
3716.4
3050
3756.9
3240
3464.3
2406

m=5
1956.7
4152
1917
4957
1890.6
5321

m=10
1500.9
4322
1405
5732
1400
5393

m=2
4739.4
5645
4810.5
5340
4241.5
2780

m=5
2409.9
7949
2358.3
10002
2365.2
8339

m=10
1786.4
9154
1650.3
10936
1650.2
10381

m=20
1445.8
8464
1417.3
12134
1415.8
5773
11

150 Cities
One Chromosome
Iterations
Two-Part Chromosome
Iterations
Modified Two-Part Chromosome
Iterations

m=2
5817.6
7732
5679.2
7205
5020.4
3272

m=5
2831.2
12154
2730.8
14810
2671.7
11777

m=10
1956.4
14269
1963.4
17879
1808
16133

m=20
1493.6
13544
1454
19096
1451.9
16448

5. Conclusion
The current work improved some of the GAs shortcomings including computation time
and falling into a local minimum. The modified GA algorithm with 2-opt mutation enhanced the
results of the min/max MTSP problem, when compared to the standard GA algorithm practices.
Overall, the longest distance and the number of iterations decreased. For further studies,
improvements may be found by exploring more efficient local search methods, trying seeded
initial solutions, and different crossover and mutation methods.

12

6. References
[1] Tolga Bektas. The Multiple traveling salesman problem: an overview of formulations and
solution procedures, Omega 34, The International Journal of Management Science, 2006,
pp. 209-219.
[2] Zhong Yu, Liang Jinhai, Gu Guochang, Zhang Rubo, Yang Haiyan, An Implementation of
Evolutionary Computation for Path Planning of Cooperative Mobile Robots , Proceeding of
the 4th World Congress on Intelligent Control and Automation, 2002, pp. 1798-1802.
[3] Donald Sofge, Alan Schultz, and Kenneth De Jong, Evolutionary Computational
Approaches to Solving the Multiple Traveling Salesman Problem Using a Neighborhood
Attractor Schema, EvoWorkshops, 2002, pp 153-162.
[4] Osman Parlaktuna, Aydm Sipahioglu, Ahmet Yazici, A VRP-Based Route Planning for a
Mobile Robot Group, Turl J Elec Engin Vol 15, 2007, pp. 187-197.
[5] Mehmed Kantardzic, Data Mining: Concepts, Models, Methods, and Algorithms, 2nd
Editions, 2011 pp. 385-401.
[6] Arthur E. Carter, Cliff T. Ragsdale, A New Approach to solving the multiple traveling
salesperson problem using genetic algorithms, European Journal of Operational Research
175, 2006, pp. 246-257.
[7] Shih-Hsin Chen, Mu-Chung Chen, Operators of the Two-Part Encoding Genetic Algorithm
in Solving the Multiple Traveling Salesmen Problem, Conference on Technologies and
Applications of Artificial Intelligence, 2011, pp. 331-335.
[8] L. Tang, J. Liu, A. Rong, Z. Yang, A Multiple Traveling Salesmens Problem Model for Hot
Rolling Schedule in Shanghai Baoshan Iron and Steel Complex, European Journal of
Operational Research 124(2), 2000, pp. 267-282.
[9] C. Malmborg, A Genetic Algorithm for Service Level Based Vehicle Scheduling,
European Joural of Operational Research 93 (1), 1996, 972-989.

13

[10] Wei Zhou, Yuanzong Li, An Improved Genetic Algorithm for Multiple Traveling
Salesman Problem, Informatics in Control, Automation, and Robotics, 2010, pp.493-496.
[11] Milan Djordjevic, Andrej Brodnik, Quantitative Analysis of Separate and Combined
Performance of Local Searcher and Genetic Algorithm, 33rd International Conference of
Information Technology Interfaces, 2011, pp. 515-520.
[12] Peng Gang, Ichiro Limra, Shigeru Nakayama, A Multiple Heuristic Search Algorithm for
Solving the Traveling Salesman Problem, Parallel and Distributed Computing Applications
and Technologies, 2003, pp. 779-783.

14

Você também pode gostar