See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3949438

Solving traveling salesman problems by combining global and localsearch

mechanisms

Conference Paper · February 2002

DOI: 10.1109/CEC.2002.1004429 · Source: IEEE Xplore

CITATIONS READS
52 184

3 authors, including:

Huai-Kuang Tsai Jinn-Moon Yang

Academia Sinica National Chiao Tung University
89 PUBLICATIONS 1,000 CITATIONS 240 PUBLICATIONS 5,086 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Huai-Kuang Tsai on 01 June 2014.

The user has requested enhancement of the downloaded file.

 Solving Traveling Salesman Problems by Combining Global and Local Search
Mechanisms
Huai-Kuang Tsai1, Jinn-Moon Yang2, and Cheng-Yan Kao1
Dept. of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan1
Dept. of Biological Science and Technology & Institute of Bioinformatics, National Chiao Tung University, Hsinchu, Taiwan2
[email protected]
Abstract - In this paper an evolutionary algorithm for the strongly related implementations of evolutionary
traveling salesman problem is proposed. The key idea is to algorithms: genetic algorithms [14], evolution strategies
enhance the ability of exploration and exploitation by [15], and evolutionary programming [16]. These three types
incorporating global search with local search. A new local of standard evolutionary algorithms are not very efficient
search, called neighbor-join (NJ) operator, is proposed to
for some specific application domains. Thus, some
improve the solution quality of the edge assembly crossover
(EAX) considered as a global search mechanism in this paper. modifications [17]-[20] have been proposed to improve
Our method is applied to 15 well-known traveling salesman solution quality and to speed up convergence. A popular
problems whose numbers of cities are from 101 to 3038 cities. approach is to combine evolutionary algorithms with local
The experimental results indicate that the neighbor-join search approaches [18], [21]. Such a hybrid approach may
operator is very competitive with the related operators possess both the global optimality of the genetic algorithms
surveyed in this paper. Incorporating the NJ into the EAX and the convergence of the local searches.
significantly outperforms the method incorporating 2-opt into
the EAX for some hard problems. For each test instance the In this paper, an evolutionary algorithm combining the
average value of solution quality stays within 0.03% from the edge assembly crossover (EAX) [1] and a new operator,
optimum. For the notorious hard problem att532, it is able to called the neighbor-join (NJ) operator, is proposed for
find the optimum solution 23 times in 30 independent runs. solving TSPs. The NJ is regarded as a local search strategy,
while the EAX is treated as a global search strategy. The basic
I. INTRODUCTION idea of the proposed approach is to enhance the ability of
The traveling salesman problem (TSP) is a well-known exploration and exploitation of the search space by the
NP-hard optimization problem, in which we require to cooperation of global and local searches. The EAX
determine the shortest closed route passing through a set of proposed by Nagata [1] was considered a powerful
n cities under the condition that each city is visited exactly crossover operator [22]. We design the NJ operator that is
once. Many problems in science, engineering, and able to improve the solution quality of EAX. According to
bioinformatics fields, such as flexible manufacturing the experimental results, the proposed approach seems to be
systems, routing as well as scheduling problems, physical able to balance exploration and exploitation.
mapping problems [2], genome rearrangement [3], and
The NJ operator reconnects cities by considering
phylogenetic tree construction [4] can be formulated as a
neighbor information and referring to the knowledge taken
TSP.
from other individuals in the population. This operator can
A large number of approaches have been devoted to be viewed as a mixture of local search, mutation and
solve the TSP. These approaches can be roughly divided recombination. The NJ operator mutates an individual itself
into local (heuristic) search and global search approaches. (mutation), but the selection of reconnecting cities depends
In general, the local search approaches, such as 2-opt, 3-opt on both the geometric neighborhood information (local
[5] and Lin-Kernigan [6] are efficient and fast convergence, search) and the edges from other individuals in the
but they may get struck at local minima. On the other hand, population (recombination). Experimental results indicate
the global search approaches, such as simulated annealing that the NJ operator is competitive with other genetic
[7], Hopfield neural networks [8], and evolutionary operators proposed for the TSPs. The average value of
algorithms [1], [9]-[13] have been proposed to reduce the solution quality stays within only 0.1% from the optimum
ill effect of these local search methods, but they often when the numbers of cities are less than 318. At the same
converge more slowly than local search approaches. time, incorporating the NJ into the EAX outperforms the
Therefore, incorporating heuristics into global search method incorporating 2-opt into the EAX.
methods has been used to tradeoff between the convergence
The rest of this paper is organized as follows. Section 2
speed and the solution quality.
introduces the evolutionary nature of the proposed
An evolutionary algorithm is based on the ideas approach. Section 3 provides the experimental results on
borrowed from genetics and natural selection. It is a some benchmark problems and comparisons of our
generally adaptable concept for problem solving, especially approach with several methods. Some characteristics of our
well suited for solving difficult optimization problems, approach are also discussed. Concluding comments are
where traditional optimization methods are less efficient. drawn in Section 4.
There are about three main independently developed but

0-7803-7282-4/02/$10.00 ©2002 IEEE

 In it ia liz e p o p u la t io n S w it h N s o lu t io n s information can be referred [1]. Two individuals, denoted
as A and B, were selected as parents. EAX first merges A
i = 0 and B into a single graph denoted G (Fig. 2). EAX travels
G to generate many AB-cycles by alternately picking edges
from parents A and B (Fig. 3). According to the heuristic
R a n d o m ly s e le c t t h e p a irin g in d iv id u a l S j
and random selection rules, some of AB-cycles are selected
I i = E A X (S i , S j, L ) to generate a quasi solution. Since edges in these selected
AB-cycles are either from A or B, those edges exist in A are
C i = N J (I i , L )
removed from A and those edges that exist in B are added
No to A.
i = i + 1 i < N The modified A then contains some disjointed subtours
Ye s (Fig. 4). The EAX uses a greedy method to merge these
R e p la c e S i w it h C i f o r e a c h i
distinct subtours into a valid solution (Fig. 5). This solution
is returned if the fitness of this solution is better than its
No
S a t is fy t e r m in a l
parents. Otherwise this procedure is repeated until a
solution is better than A and B or L children are produced.
Ye s

O u t p u t t h e b e s t s o lu t io n

Fig. 1. The outline of the proposed evolutionary algorithm

II. APPROACH
In this section the details of the proposed genetic
algorithm for TSP are presented. Our algorithm consists of (a). SolutionA (b). Solution B (c). Graph G
the edge assembly crossover (EAX) for global search and a Fig. 2. A, B, and the graph G. G is obtained by overlapping A and B.
new genetic operator for local search, called neighbor-join
(NJ). The EAX and the NJ are used to balance exploration
and exploitation in the search space.
Fig. 1 shows the main steps of the proposed
evolutionary algorithm. N solutions are generated as the
initial population. Each solution is represented as a
permutation from 1 to n, where n is the number of cities of (a) (b) (c). (d)
a TSP. After evaluating the fitness, each individual (Si) in
the population sequentially applied the following steps to Fig. 3. An example of the division of undirected edges into AB-cycles on
G. (a), (b), and (c) are effective AB-cycles, while AB-cycles in (d)
produce a child: Si randomly selects another individual Sj are ineffective.
and these two individuals produce an intermediate offspring
(Ii) by conducting the EAX. Then the NJ is executed L
times (L is the local search length) to generate a child Ci by
refining from the offspring Ii. Each solution in the
population executes above steps to generate its child and
these N children become the new population of the next
generation. (a). (b). (c). (d).

Our algorithm terminated when one of following Fig. 4. Four examples of EAX individually apply some AB-cycles on A.
criteria is satisfied: 1) all individuals of a population are the (a), (b), and (c) are the results of applying AB-cycle shown in Fig.
3(a), 3(b), and 3(c), respectively, and (d) is the result of applying
same, or 2) all of the children generated in continuous ten AB-cycles shown in Fig. 4(a) and 4(b) on A.
generations are worse than their parents. In the following
subsections, the EAX and the NJ are described.

A. EAX
The EAX proposed by Nagata [1] was considered a
powerful crossover operator. The EAX has two important
(a). (b). (c). (d).
features: preserving parents’ edges using a novel approach
and adding new edges by applying a greedy method, Fig. 5. After Modification, all examples in Fig. 5 are
analogous to a minimal spanning tree. The EAX is individually modified to valid solutions.
considered as the global search strategy in our proposed
algorithm.
The EAX is briefly described here and the detail

0-7803-7282-4/02/$10.00 ©2002 IEEE

 to the analysis of the optimal tour of att532, we find that
NJ (Ii, L) : Given an intermediate individual Ii and local most of the links in the optimal tour of att532 are the
search length L neighbor cities of each city.
Let Ii’ = Ii Fig. 6 outlines the steps of the NJ. By given the input of
repeat L times an individual Ii and the search length L, the NJ applies L
{ modifications from the start solution Ii’= Ii. In each
randomly select a city c from Ii’ modification, a city c is randomly selected from Ii’. With
if (rand() > 0.5) equal probability, a city c’ is randomly selected either from
Randomly select another individual from current the geometric nearest three neighbors of c or from an
population. Select the city c’ connecting to c in the individual, which is randomly selected from the population.
selected individual. If the city c’ does not connect to the city c in Ii’, to align c
else randomly select c’ from the geometric nearest and c’ together generates four possible types shown in Fig.
three cities of c. 7. The NJ generates four candidates by sequentially
if the city c’ doesn’t connect to c in Ii’ executing each type once. The one with lowest fitness of
a. To align c’ and c together to generate four these four candidates and Ii’ are selected as the parent of
candidates (see Fig. 7). next loop. Above steps are executed L times.
b. Select the one with lowest fitness from these
In the four candidates Types I and II are the simple
four candidates and Ii’ to replace Ii’. invert operation to align c and c’ together. For Types III
} and IV a greedy method is applied to merge two disjoint
return Ii’


subtours into a valid solution. The greedy method works as
Fig. 6. The outline of the neighbor-join operator follows: Let vi represent a city, (vi, vj), i j, represents an
edge, and w(vi, vj) be the edge length of (vi, vj). At the same
time, let (vr, vr+1) and (vs, vs+1) be the edges of the subtour
Ii’ 0 1 2 3 4 5 6 7 8 9 Gr and the subtour Gs, respectively. We find a pair of edges
c c’ (vr, vr+1) and (vs, vs+1) to connect these two subtours, Gr and
Gs, into a legal tour by maximizing the value of the
TypeI 0 1 2 3 7 6 5 4 8 9 following equation
c c’
w(vr , vr +1 ) + w(vs , vs+1 ) − w(vr , vs+1 ) − w(vs , vr +1 ) ∀r , s; r ∈ Gr and s ∈ Gs .

TypeII 0 1 2 6 5 4 3 7 8 9
The new edges, (vr, vs+1) and (vs, vr+1), are inserted to
c c’
replace the original edges, (vr, vr+1) and (vs, vs+1), to form
After modify
the new solution.
0 1 2 4 3 7 6 5 8 9
TypeIII 0 1 2 3 4 5 6 7 8 9
c c’
c c’ C. AN EXAMPLE OF THE NEIGHBOR-JOIN
After modify
Let’s illustrate a single iteration of the NJ on the
TypeIV 0 1 2 3 4 5 6 7 8 9
0 1 5 4 6 2 3 7 8 9 following example. Assume that the current individual Si is
c c’
c c’ Ii’ = (0,1,2,3,4,5,6,7,8,9)
Fig. 7. Four possible types are considered to connect city c and c’ together.
and the current city c is 3. If rand() > 0.5, another
The results of Type I and Type II are obtained via the simple invert
individual, (0,3,7,2,9,8,4,1,5,6), is randomly selected from
operation. The results of Type III and Type IV are constructed by
the current population. Since the city 7 connect the city 3 in
applying a greedy method (see text) from two disjoint subtours.
this individual and does not align next to the city 3 in
individual Ii’, the cities 3 and 7 are decided to be connected.
B. NEIGHBOR-JOIN OPERATOR Four types of candidate are considered (see Fig. 7) and the
one with lowest fitness function value is accepted. If type III
The neighbor-join (NJ) operator constructs a new
is the lowest one and the gain of exchanging edge-pair (v2,
solution by stealing edges from other individuals in the
v8)- (v4, v5) by edge-pair (v2, v4) - (v5, v8) is maximum, Ii’ is
population or by considering the geometric information.
updated as (0,1,2,4,3,7,6,5,8,9) if the fitness function value
Although the NJ is applied only on the single solution, the
is better than the original Ii’.
offspring is generated considering both the neighborhood
information and knowledge from other individuals. Thus, Again, randomly assign the new city c as 5. If one, c, of
the NJ operator is a genetic operator combining with the the geometric three nearest cities is selected, assume, city 4
characteristics of local search, mutation, and recombination. is determined to connect to city 5. Four candidates are
considered:
The NJ is inspired by the inver-over mutation [11] and
by analyzing the TSP search space [20]. The main Type I : (0,1,2,4,5,6,7,3,8,9)
difference between the inver-over mutation and the other Type II : (0,1,2,6,7,3,4,5,8,9)
mutations is that it inherits edges both from parent and Type III : (0,1,2,8,9,0) (4,3,7,6,5,4)
from other individuals in the current population. According Type IV : (0,1,2,4,5,8,9,0) (3,7,6,3)

0-7803-7282-4/02/$10.00 ©2002 IEEE

If type I is the one with lowest fitness function value, Ii’ is TABLE I.
updated as (0,1,2,4,5,6,7,3,8,9) if the fitness function value Comparisons among neighbor-join, 2-opt and city-swap genetic operator
is better. Above steps are repeated for L times. on TSP problems taken from TSPLIB [23] based on the error rate, the
number of optimal solutions found vs the number of trails, average
III. EXPERIMENTAL RESULTS generation, and average CPU time in 50 trails. In all cases, the error rate of
In this section, some well-known TSPs [23] are used to NJ is always best (only 0.1% above the optimal tour length) and NJ is
illustrate the robustness of our proposed approach. First, quicker than other operators for larger problems. Here the error (%) is
two important parameters, search length L and population defined as averge − optimum .
optimum
size N, are discussed. Next, we compare the NJ with 2-opt Problem Method Error rate Optimal Gen(CPU)
and city swap to verify the effectiveness of the NJ. Finally, (%) times
NJ 0.012 47/50 2.14(0.12)
the proposed algorithm is compared with some other att48 2-opt 0. 490 5/50 19.00(0.266)
methods. The proposed algorithm was implemented in C++ cityswap 1.673 1/50 12.00(0.180)
NJ 0. 097 29/50 6.46(0.20)
and run on a Pentium III 500MHz personal computer with eil51 2-opt 1.409 1/50 19.76(0.166)
cityswap 2.859 0/50 11.16 (0.180)
single processor. NJ 0. 294 16/50 11.64(0.66)
eil76 2-opt 4.368 0/50 35.86(0.500)
cityswap 11.66 0/50 20.00 (0.520)
A. PARAMETERS SETTINGS NJ 0. 550 7/50 17.78(1.44)
eil101 2-opt 7.043 0/50 55.20(1.033)
cityswap 21.93 0/50 35.16 (1.180)
The population size (N) and the search length (L) are NJ 0. 311 1/50 29.22(9.28)
main parameters in our algorithm. To decide the parameter kroA200 2-opt 12.85 0/50 204.0(8.500)
cityswap 93.32 0/50 220.9 (13.06)
values, we have tested various values of these two NJ 0. 285 5/50 35.80(15.40)
pr264 2-opt 26.93 0/50 286.5(17.16)
parameters on 15 TSP problems selected from TSPLIB[23]. cityswap 159.4 0/50 405.8 (30.86)
NJ 0. 894 0/50 47.20(30.06)
Fig. 8 shows the relationship between the solution lin318 2-opt 29.06 0/50 358.9(30.50)
cityswap 173.9 0/50 493.8 (47.08)
quality and the convergence time of our algorithm tested on
problem att532 with various population sizes (N).
According to the experiments we set the population size To understand the influences of the search lengths of the
roughly equal to the number of cities of a problem for the EAX and the NJ operators, we observed the performance
tradeoff between the solution quality and the convergence with various search lengths (L). Fig. 9 shows the
time. performance of the purposed algorithm on 15 TSP
problems when the values of L are set to 5, 20, and 50,
0.00030
respectively. Each problem was tested 30 distinct trails. For
300
each problem the proposed approach has the worst
250 0.00025
performance when L is 5 and the improvement of the
Error rate

0.00020
200
solution quality is not significant when L exceeds 20. In
0.00015
Time

150 practice, the longer L is, more time will spend. Therefore, L
0.00010
100 is set to 20 in this paper.
50
0.00005
0 Time
0
100 200 300 400 500 600 700 800 900 1000 Error
Population size rate B. THE PERFORMANCE OF NJ
Fig. 8. Relationship among population size, error rate, and Table I summarizes the results of the NJ operator, 2-opt,
convergence time of our algorithm tested on the att532 problem. and city swap on some benchmark problems. Each problem
was tested 50 independent runs. In the table, the “Error

rate (%)” is the relative error of the average of best value

to the optimum in percentage; The column, “Opt/trail”
denotes the number of finding the optimal solution in 50


trails; and “Gen (CPU)” denotes the values of the average
Optimal times


Length = 5 generation and average CPU time required for finding the
Length = 20 best tours, respectively. By the observation, the solution


Length = 50
qualities of the NJ operator outperform the 2-opt and city
 swap. For all test instances the average value of solution


quality stays within only 0.1% from the optimum. For the
problem lin318, although the NJ is not able to find the
Eil101
KroA200
Lin318
Pcb442
Att532
U574
Rat575
U724
Rat783
Vm1084
Pcb1173
U1432
Vm1748
Pr2392
Pcb3038

optimum solution, it can still find near-optimal tours in

Problems
each run and the average tour length that was only 0.894%
Fig. 9. Comparisons of the proposed algorithm on 15 TSP problems
above the optimum tour length. By comparing the famous
with different search lengths 5, 20, and 50 based on the number of finding 2-opt, these results show that the NJ is more robust than 2-
the optimal solution in 30 trails. The performance was worst when L is set opt for these testing TSPs.
to 5, and it has similar performance when the values of L are 20 and 50.

0-7803-7282-4/02/$10.00 ©2002 IEEE

 C. THE IMPROVEMENT OF APPLYING NJ TABLE III

Here we analyzed the influences of our method by Comparisons of the proposed algorithm (denoted as HGA) with other
cooperating different heuristics, such as 2-opt and NJ methods, including NGA [24], ACS [10], DGA [9], and EGA[1], on five
mutation, into the EAX operators. Table II shows the TSP problems based on the best tour length, average tour length, and the
results of our methods with applying different operators for standard deviation. “N/A” represents not available in original papers.
15 TSP problems. According to the results, the methods Problems/ σ
combining 2-opt into EAX is limited for improving the Methods Best (Error %) Average (Error %)
(optimum) n
solution quality. On the other hand, the method combining ACS N/A N/A N/A
the NJ into EAX significantly promotes the solution Lin318
DGA 42029(0.000) 42033.44(0.011) 1.35
qualities for all test problems. Although the NJ cannot be NGA 42029(0.000) 42029.00(0.000) 0.00
(42029)
EGA 42029(0.000) 42041.23(0.011) 1.81
proved to be complementary with EAX, incorporating the HGA 42029(0.000) 42029.00(0.000) 0.00
NJ with the EAX indeed improves the solution qualities. ACS N/A N/A N/A
The key novelty of the present work is the seamless ability DGA 50778(0.000) 50778(0.000) 0.00
pcb442
of integrating global and local search mechanism through (50778)
NGA 50778(0.000) 50778(0.000) 0.00
incorporation of a number of genetic operators each having EGA 50778(0.000) 50778(0.000) 0.00
HGA 50778(0.000) 50778(0.000) 0.00
unique search mechanisms.
ACS 27693(0.000) 27718.20(0.112) 0.72
DGA 27686(0.000) 27697.58(0.042) 0.48
att532
D. COMPARISONS WITH OTHER METHODS NGA 27686(0.000) 27695.61(0.035) 0.71
(27686)
EGA 27686(0.000) 27696.33(0.037) 0.62
Following the discussions of the proposed algorithm HGA 27686(0.000) 27688.49(0.009) 0.00
(denoted as HGA) in the previous subsection, we compare ACS 8818(0.136) 8837.90(0.362) 0.19
DGA 8806(0.000) 8806.00(0.000) 0.00
the results of the proposed approach with the best-known rat783
NGA 8806(0.000) 8806.00(0.000) 0.00
results of four methods based on our best knowledge. Five (8806)
EGA 8806(0.000) 8806.00(0.000) 0.00
TSP benchmark problems, including lin318, pcb442, att532, HGA 8806(0.000) 8806.00(0.000) 0.00
rat783, and pcb3038, shown in Table III are used because ACS N/A N/A N/A
DGA 137705(0.008) 137760.55(0.048) 4.28
they have been widely used to compare the performance pcb3038
NGA 137695(0.001) 137765.02(0.052) 4.55
among algorithms. Our proposed algorithm was executed (137694)
EGA 137694(0.000) 137703.77(0.007) 0.93
30 trials for each problem. HGA 137694(0.000) 137700.19(0.004) 0.00

TABLE II
Comparisons of our method with applying different local search methods. Table III summarizes the results of our method and four
These methods are tested on 15 TSP problems based on the average CPU
other approaches, including nature crossover genetic
time (Time), the number of trails that optimal solutions found (Opt times),
and average solution qualities in 30 trails (Error). Here the error (%) is algorithm (NGA) [24], ant colony system (ACS) [10],
defined as average − optimum . distance-preserving crossover genetic algorithm (DGA) [9],
optimum
and EAX genetic algorithm (EGA) [1]. NGA integrated
All of three methods use the same parameters, including search length nature crossover and LK local search [6]; ACS is an ant
(L), population size (N), EAX crossover and stop criteria. The only
difference is the use of local search.
colony system cooperated with 3-opt operator; DGA
combined the distance-preserving crossover and 3-opt; and
Local
search
No local search NJ 2-opt EGA used the EAX crossover. These four approaches
Time Opt Error Time Opt Error Time Opt Error perform well on these test problems according to our
problem surveys. The results of first three methods were directly
(Sec.) times (%) (Sec.) times (%) (Sec.) times (%)
eil101 1.03 30 0.0000 1.50 30 0.0000 1.22 30 0.0000 summarized from original papers [9][10][24]. The best tour
kroA200 3.24 30 0.0000 4.82 30 0.0000 3.97 30 0.0000 length, average tour length, and standard deviation of trails
lin318 14.54 26 0.0291 19.1 30 0.0000 15.65 27 0.0191 are used to measure the performance of comparative
pcb442 39.91 30 0.0000 52.26 30 0.0000 43.57 30 0.0000 methods. The values in parentheses of the best and the
att532 97.93 7 0.0373 128.2 24 0.0090 105.4 11 0.0312 average tour length represent the percentages of error
u574 64.65 30 0.0000 89.47 30 0.0000 76.89 30 0.0000 sol − optimum
defined as optimum , where sol is the experimental
rat575 128.73 21 0.0043 172.9 28 0.0009 147.5 23 0.0031
u724 257.60 18 0.0122 335.4 27 0.0019 279.3 16 0.0059 value and optimum is the optimum of a TSP problem. The
rat783 σ
375.45 30 0.0000 487.3 30 0.0000 396.7 30 0.0000 column “ ” shows the group standard deviation of trails


n
vm1084 810.12 18 0.0277 1045 21 0.0227 854.9 20 0.0099
pcb1173 1089 24 0.0050 1401 29 0.0005 1171 27 0.0005 where is the standard derivation and N is the number of
u1432 2034 16 0.0145 2664 26 0.0080 2160 24 0.0038 cities.
vm1748 4295 19 0.0189 5663 27 0.0074 4633 21 0.0125 Table III shows that our algorithm performs better and
pr2392 9989 21 0.0049 13172 30 0.0000 10750 28 0.0009 more robustly than comparative methods for testing
pcb3038 22341 24 0.0071 29813 28 0.0045 24296 25 0.0006 problems. For each problem the proposed algorithm can
find the best tour in almost each trial and the error rate is
only 0.01% away from the optimal.

0-7803-7282-4/02/$10.00 ©2002 IEEE

 IV. CONCLUSION [14]. Goldberg, D. E. (1989) Genetic algorithms in search, optimization
& machine learning. Reading. MA: Addison-Wesley.
This study demonstrates that combining suitable global [15]. Baeck, T. (1996) Evolutionary Algorithms in Theory and Practice.
and local search is an encouraging approach for solving Oxford University Press, New York, USA.
TSPs. Experimental results verify that the proposed [16]. Fogel, D. B. (1995) Evolutionary Computation: Toward a New
approach is able to generate efficient solutions for TSPs. Philosophy of Machine Intelligent. NJ:IEEE Press, Piscataway.
The proposed neighbor-join operator, combining with local [17]. Xiao, J., Michalewicz, Z., Zhang, L., and Trojanowski, K. (1997)
search, mutation, and recombination together, outperforms “Adaptive evolutionary planner/navigator for mobile robots.” IEEE
other genetic operators surveyed in this paper. Moreover, Trans. On Evolutionary Computation, 1(1): pp.18-28.
our approach can find out the near-optimal solutions with [18]. Hart, W. E. (1994) Adaptive global optimization with local search.
only about 0.03% above the optimum. We believe that the PhD thesis, University of California, San Diego.
flexibility and robustness of our approach make it an [19]. Saravanan, N. and Fogel, D. B. (1997) “Multi-operator evolutionary
effective tool for TSPs. programming: A preliminary on function optimization.” In P. J.
Angeline et al. editor, Proc. 6th Annu. Conf. On Evolutionary
Programming (Lecture Notes in Computer Science, vol. 1213),
pp.215-222.
References
[20]. Back, T., Hammel, U., and Schwefel, H. -P. (1997) “Evolution
[1]. Nagata, Y., Kobayashi. S. (1997) “Edge assembly crossover: A high- computation: Comments on the history and current states” IEEE.
power genetic algorithm for the traveling salesman problem.” In T. Transactions on Evolutionary Computation. pp. 3-17.
Back, editor, Proc. of the 7th Int’l. Conf. on GAs, pp. 450-457. Morgan
Kaufmann. [21]. Yen, J., Yip, J. C. and Pao, Y. H. (1998) “Combinatorial optimization
with use of guided evolutionary simulated annealing.” IEEE. Trans. On
[2]. Alizadeh, F., Karp, R.M., Newberg, L.A., and Weisser, D.K. (1993) Systems, Mans, and Cybernetics –Part B, 28(2): pp.173-191.
“Physical mapping of chromosomes: a combinatorial problem in
molecular biology.” In Proc. 4th ACM-SIAM Symposium on Discrete [22]. Watson, J., Ross, C., Eisele, V., Denton, J., Bins, J., Guerra, C.,
Algorithms (SODA). pp. 371-381. Whitely, D., and Howe, A. (1998) “The traveling Salesrep Problem,
edge assembly crossover, and 2-opt.” In Parallel Problem Solving
[3]. D. Sankoff and M. Blanchette. (1997) “The median problem for from Nature V, A.E. Eiben et al, eds. Springer-Verlag. pp.823-832.
breakpoints in comparative genomics.” Computing
and Combinatorics,Proceedings of COCOON `97. Lecture Notes in [23]. Reinelt, G. (1991) “TSPLIB- A Traveling salesman problem library.”
Computer Science 1276, Springer Verlag, New York, pp. 251-263. ORSA Journal on Computing, Vol.3, No.4, pp.376-384.
[4]. Korostensky, C. and Gonnet, G. H. (2000) “Using traveling [24]. J. Soonchul and M. Byung-Ro (2000) “The nature crossover for the
salesman problem algorithms for evolutionary tree construction. 2D Euclidean TSP.” in Proceedings of the Genetic and Evolutionary
“ Bioinformatics Vol. 16 no. 7, pp. 619-627. Computation Conference (GECCO), pp. 1003-1010.
[5]. Lin, S. (1965) “Computer solutions of the traveling salesman
problem.“ Bell Syst., J., vol. 23, pp. 2245-2269.
[6]. Lin, S., and Kernighan B. (1973) “An effective heuristic algorithms for
the traveling salesman problem.” Operations Research Vol.21, pp.498-
516.
[7]. KirKpatrick, S., Gelatt C. D., and Vecchi M. P. (1983) “Optimization by
simulated annealing.“ Sci, vol. 200, pp. 671-680.
[8]. Hopfield, J. J. and Tank, D. W. (1985) “Neural computation of
decisions in optimization problems. “ Biological Cybernetics, vol.
52, pp. 141-152.
[9]. Freisleben, B. and Merz, P. (1996) “New genetic local search
operators for the traveling salesman problem.“ In Proc. PPSN IV-
4th Int. Conf Parallel Problem Solving from Nature. Berlin,
Germany: Springer-Verlag, pp. 890-899.
[10]. Dorigo, M., and Gambardella, L. M. (1997) “Ant colony system: A
cooperative learning approach to the traveling salesman
problem.“ IEEE. Transaction on Evolutionary Computation, vol.1,
no.1. pp. 53-66.
[11]. Tao, G., and Michalewicz, Z. (1998) ”Inver-over Operator for the
TSP. ” Proceedings of the 5th Parallel Problem Solving from
Nature, September 27-30, 1998, Lecture Notes in Computer Science,
pp.803-812.
[12]. Mulhem, M., and Maghrabi, T. (1998) “Efficient convex-elastic net
algorithm to solve the Euclidean traveling salesman problem.”
Systems, Man and Cybernetics, Part B, IEEE Transactions, Vol.284,
pp.618 –620.
[13]. Zhenya, H., Chengjian, W., Bingyao, J., Wenjiang, P., and Luxi. Y.
(1999) “A new population-based incremental learning method for
the traveling salesman problem.” Proceedings of the 1999 Congress
on Evolutionary Computation-CEC 99. pp.1150-1156

0-7803-7282-4/02/$10.00 ©2002 IEEE

View publication stats