Proceedings of the 19th World Congress

The International Federation of Automatic Control

Cape Town, South Africa. August 24-29, 2014

Colored Traveling Salesman Problem and Solution


Jun Li1, 2, Qirui Sun1, 2, MengChu Zhou3, Xiaolong Yu1, 2, and Xianzhong Dai1, 2

1. Key Laboratory of Measurement and Control of CSE, Ministry of Education, Nanjing 210096, China
2. School of Automation, Southeast University, Nanjing 210096 China (e-mail: {j.li, xzdai}@seu.edu.cn; {sunqr,
xly19900107}@163.com)
3. New Jersey Institute of Technology, Newark, NJ 07102 USA (e-mail: [email protected])

Abstract: Multiple Traveling Salesman Problem (MTSP) is an important combinatorial optimization

problem. However, it is applicable to only the cases in which multiple executing individuals (traveling
salesman) share the common workspace (city set). It cannot be used to handle many multi-machine
engineering systems where multiple machines’ workspaces are not the same and partially overlap with
each other. This paper proposes and formulates a new MTSP called colored traveling salesman problem
(CTSP). Each of its salesmen is assigned a private city set and all salesmen share a public city set. Every
set of cities is colored differently. To solve CTSP, we present two improved genetic algorithms (GA) by
combining the classic one with a greedy algorithm and hill-climbing one to achieve better performance.
Finally, the algorithms are applied and compared through a case study. The result shows that the hill-
climbing GA enjoys the best performance among the investigated ones.
Keywords: TSP, MTSP, Modelling, Genetic Algorithm, Greedy Algorithm, Hill-climbing Algorithm

salesman, i.e., each city can be visited by any salesman.
1. INTRODUCTION
However, not all executing individuals have the same
A multiple traveling salesman problem (MTSP) generalized workspace in some application problems. Take the
from a traveling salesman problem (TSP) is a well-known scheduling of a multi-machine engineering system (MES) as
combinatorial optimization problem. It aims to determine a an example. The workspaces of individual machines are not
family of tours with minimal total cost for multiple salesmen the same but overlap partially with each other. Thus, each
to visit each city exactly once within a given set and machine has to perform not only the operations
eventually return to the home city. MTSP and TSP arise in a independently in its private workspace, but also complete all
variety of applications that require addressing scheduling, the operations with other machine(s) together in the
planning, routing, and/or sequencing issues. Application overlapped workspace. A typical MES, i.e., a dual-bridge
examples of TSP in machine scheduling and sequencing, and waterjet cutting machine tool, is illustrated in Fig. 1.
vehicle routing can be founded in (Gutin and Punnen, 2002).
Other work reports additional applications in circuit wiring
(Hirogaki, et al., 2005) and in statistical data analysis
including ordering and clustering objects, e.g., gene ordering
in (Ray, et al., 2007) and protein clustering in (Johnson and
Liu, 2007). The most comprehensive survey on the
applications of MTSP is given in (Bektas, 2006). Carter and
Ragsdale (2006) stress its use in pre-print insert
advertisement scheduling. A similar application in hot rolling
scheduling is reported in (Tang, et al., 2000). Autonomous
robot or vehicle motion planning (Basu, et al., 2000, and
Ryan, et al., 1998) represents other types of its applications.
Saleh and Chelouah (2004) apply it to satellite surveying Fig. 1. Dual-bridge Waterjet Cutting Machine Tool
system design. Toth and Vigo (2002) investigate a vehicle
routing problem as the generalization of TSP. Cheong and It consists of two independent bridge systems. Their cutting
White (2012) have investigated how to dynamically areas have an overlapped section, i.e., the marked area on the
determine a tour for TSP based on real-time traffic workbench with a red box, so as to prevent the presence of
congestion data. cutting dead zone. Thus, the overlapped area allows both
bridges to enter and the two areas out of it are their exclusive
In essence, MTSP is an abstraction of the practical problems cutting areas.
in which multiple executing individuals (traveling salesmen)
are involved and they share the common workspace (city set). Due to the partially overlapped workspaces, a scheduling
In other words, all the cities of MTSP are identical for each method for MTSP cannot be simply used to schedule MES.

978-3-902823-62-5/2014 © IFAC 9575

19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

On the other hand, the basic elements of such a problem, i.e., Fig. 2. Example of CTSP 
individuals, operations, and workspaces, are still similar to
The nodes in the areas V1 , V2 , and V3 represent the private
the salesmen, city visits and city set of TSP, respectively. The
difference lies in that each individual (salesman) of the cities of salesmen 1, 2, and 3. They have the only shared city
former not only has a private workspace (city set) but also set V0 .  
shares a common workspace with others. To distinguish the
different cities, we define a new multiple traveling salesman
2.2 0-1 Integer Programming Model
problem by coloring the cities, called Colored TSP (CTSP).
CTSP frequently arises in real-life applications where some
CTSP is formulated over a complete digraph G  (V , E ) ,
closely dependent relations between the salesmen and the
cities must be obeyed when one determines a solution. It is a where the vertex set V  0,1, 2,..., n  1 corresponds to the
significant problem in theory and practice. We call the same cities and each edge in (i, j )  E ， i  j , is associated with a
problem as MTSP* by Li, et al. (2013) and present a genetic
weight ij representing a visit cost (distance) between two
algorithm (GA) solution. However, CTSP has not been
formulated in a mathematically rigorous way. This paper cities i and j . The vertex 0 represents the home city (depot).
formally defines CTSP and improves the prior method in (Li, Let V  V \ 0 . V is divided into m+1 sets, i.e., V , the 0
et al., 2013) by combining GA with Greedy Algorithm and
Hill-climbing Algorithm. public one, and Vi , the private ones of the salesmen for all
i  Z m . The objective of CTSP is to determine m
Next, CTSP is formulated in Section 2. Section 3 presents
Hamiltonian cycles or circuits on G with the least total cost
two improved GAs. Section 4 gives a case study with the
such that any vertex of each private set is visited exactly once
comparison results. The paper is concluded in Section 5.
by the specified salesman and any vertex of the public set is
visited by any salesman exactly once and eventually return to
2. DEFINITION AND FORMULATION OF CTSP
city 0. Of course, it allows each salesman to have an
exclusive home city in their private set, like the waterjet
2.1 CTSP Definition cutting example shown in Fig. 1. However, this work will not
discuss such cases.
Let n, m  Z  1, 2,3,... and m  n . CTSP aims to
Binary variable xijk =1 , i  j , i, j  V , and k  Z m , if the k-th
determine a family of tours with the minimal total cost for m
salesmen to visit n cities exactly once given public and salesman passes through edge (i, j); and otherwise, xijk  0 .
private ones, and eventually return to the home city (depot). uik  is the number of nodes visited on the k-th salesman’s tour
Let Vi , i  Z m  1, 2,..., m be the private city set assigned
from the depot up to node i. 
to the i-th salesman and U j , j  Z r , r  1 be the j-th shared
The integer programming model of CTSP is presented as
city set, and Wi be the accessible city set of the i-th salesman, follows.
i.e., the union of sets of i-th salesman’s private cities and
m n 1 n 1
shared cities.
Minimize  ij xijk (3)
k 1 i  0 j  0
The city sets meet the following constraints, given i  j :
n 1
U i  U j   ,  i, j  Z r (1) Subject to x
i 1
0 ik  1, (4) 

Wi  W j and W j  Wi ,  i, j  Z m (2) n 1

There are various cases of the intersections among the

x
i 1
i 0k =1 , k  Z m , (5)

accessible city sets in CTSP. A common one is that there is

only one common city set shared by m salesmen, as shown in  x ijk  0, (6) 
Fig. 2. i j

 x
i j
jik  0 , i  Vk , j  V \ (V0  Vk ), k  Z m , (7)
V2

 x
i j
ijl 0, (8)
V1 U0

 x
i j
jil  0 , i  Vk , j  V , i  j , k  l , l  Z m , (9)

V3
n 1 m

 x ijk 1, (10)

  j  0 k 1

9576
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

n 1 m  
 x
j  0 k 1
jik  1 , i  V , j  i , (11)

x
h
jhk   xijk , j  V0 , i, h  Vk  V0 , i  j  h ,
i
(12)

uik  u jk  n  xijk  n  1 , i, j  V , j  i, k  Z m , (13)

Equations (4) and (5) require that every salesman start from
and return to city 0, and (6) ensures that salesman k cannot
start from his own exclusive city to visit a private city of
other salesmen and (7) that another salesman is forbidden to
Fig. 3. Example of CC
visit a private city of salesman k from its own private city as
well. Equations (8) and (9) ensure that salesman l (  k ) can In Step 1, given two parents, a section of a city individual is
neither start from a private city of salesman k nor return to it. selected at random, and then its genes are swapped with those
Each city except city 0 can be visited exactly once as of another individual, thereby resulting in two new
described by (10) and (11). Equation (12) represents that a individuals as shown in Step 2. The mapping relationship of
public city can be visited by any salesman while (13) the selected sections in two city individuals is 8—3, 9—8,
prohibits the formation of any sub-tour among nodes in V 5—2, 4—7, 7—1, and 1—10. Step 3 exchanges the
\{0}. redundant genes according to the selected section, and then
finds that private cities 5, 3, 7, 1, and 6 in the left
Like MTSP, CTSP is also NP-hard. Moreover, the restriction chromosome and cities 2, 5, and 4 in the right one are
on city colours makes its solution more difficult and time- assigned to the wrong salesmen. Next, Step 4 reassigns the
consuming than that of MTSP. It is proven that the heuristics private cities to the correct salesmen and obtains two
are faster and more efficient than the exact methods in the reasonable generations.
solution of MTSP with respect to the problem size. In many
cases, however, the former cannot be guaranteed to obtain the A city mutation (CM) process in a dual-chromosome is
optimal solution and are thus applicable to solve those cases illustrated in Fig.4.
in which good-quality solutions suffice. With this in mind,  
this paper presents GA for CTSP.

3. GENETIC ALGORITHMS FOR CTSP

3.1 GA and Its Limitation

Our prior work developed a basic GA to solve CTSP (Li, et

al., 2013). It represents a solution via dual chromosomes that
are decimally coded, i.e., city and salesman chromosomes. Fig. 4. Example of city mutation (CM)
Constraints (6)-(12) as a city assignment relation are taken
First, the gene points of cities 8 and 7 are selected as
account into the dual-chromosome coding where each city
swapping ones. After swapping, the city assignment relation
gene corresponds to a right salesman gene at the same
is satisfied and the mutation is over.
position. Suppose that private cities of salesmen 1-3 are cities
1-2, 3-4, and 5-6, respectively, and the shared cities are cities The fitness function takes the value of the inverse of the total
7-10. tour cost f(x) equal to that of Eq. (3).
Li, et al., (2013) adopt the combination of Roulette Wheel 1
method and Elitist strategy as the selection operation. Three F ( x)= (14)
1+f (x)
pairs of compositions of the crossover and mutation operators
are compared and the result shows that the performance of In (Li, et al., 2013), the case study indicates that the
city crossover and city mutation (CCM) operator is the best. evolution of GA is slow and it is easy to trap in a local
A city crossover operator is a modified partially matched optimum.
crossover (PMX). Figure 3 shows a crossover process of a
dual-chromosome with a single crossover of city
chromosomes. 3.2 Greedy GA

The decision made by using Greedy Algorithm at each step

may not reach the best in the global view but the local
optimum. However, it can obtain the satisfactory solution
rapidly because it avoids the great effort needed to exhaust all
possibilities to find the optimal solution. We use it to

9577
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

optimize the individuals of the initial population generated algorithm is very strong and it is a common method used for
randomly at the first step of GA. Initial population of high the local optimum search.
quality will accelerate the population evolution of GA and
reach satisfactory solution rapidly. We name this improved GA in (Li, et al., 2013) adopts the combinatorial selection
algorithm as a greedy GA. strategy of elite reservation and roulette in (Sun, 2013). After
a certain period of evolution, it may be trapped into a local
With regard to CTSP, the criterion is defined as the shortest optimum. To escape from it, the best individual of each
distance between two cities. Namely, a city will be selected generation can be optimized by using Hill-Climbing
as the next one that the corresponding salesman will visit Algorithm. Specifically, if a better individual is obtained
once it is nearest to the current city in the visited sequence. It through hill-climbing, it replaces the original one; and
can optimize a solution by reordering its sequence. For otherwise, the original one remains in it. Note that the hill-
example as shown in Fig. 5, the randomly generated visit climbing GA adopts Greedy Algorithm to optimize the initial
sequence is 0→2→5→3→1→4→0. The sum of distances is population too.
25+50+50+45+45+25=240. It can be optimized to be
The neighbourhood point selection impacts greatly on the
sequence 0 → 3 → 4 → 5 → 1 → 2 → 0 by using the greedy
hill-climbing algorithm and the paper adopts two point
algorithm. The improved distance is 22+25+27+35+25+
swapping. Given CTSP with m( 2) salesmen, it should
25=159. Obviously, it is a better solution.
select 2m genes by this selection strategy. The fitness must
be recalculated after every time of gene swapping.
A hill-climbing GA includes the following steps:
Step 1: Determine if the i-th salesman performing the current
swapping is the m-th salesman, i.e., i =m . If so, end this hill-
climbing; otherwise, go to the next step.
Step 2: Select two city genes assigned to the i-th salesman,
from the city chromosome of a. Swap them and obtain
individual a  , and go to the next step.
Fig. 5. Distances between cities
Step 3: Determine if the value of fitness of a  is greater than
The generation process of the initial population in greedy GA that of a. If so, let a  a  ; and otherwise, give up a  and keep
is as follows. a.
Step 1: Determine if the number of individuals in the current Step 4: Let i  i  1 , and return to Step 1.
initial population is equal to the set number N or not. If it is
true, terminate the process; otherwise, go to the next step. The main procedure of hill-climbing GA is summarized in
Fig. 6.
Step 2: Generate a city sequence randomly and assign the
private cities to the specified salesman and the public cities to
all the salesmen randomly. It results in individual a.
Step 3: Reorder the city sequence of a by the shortest
distance criterion to minimize the visit cost and obtain
individual a  .

Step 4: Detect if a  has already existed in the population or

not. If so, go back to Step 2; otherwise, insert it into the
population and go back to Step 1.

3.3 Hill-Climbing GA

Hill-climbing Algorithm utilizes neighbourhood search

techniques to search, like hill-climbing, in a single direction
that the quality of a solution is possible to be improved (Lim,
et al., 2006). Starting from an existing node, it generates a
new solution with a method of neighbourhood point selection
and compares it with the value of the existing node. If the
former is larger, replaces the latter by the former; otherwise,
return the latter and set it as the maximum. Repeat the
process of climbing upward (to better solution) until the
highest point is reached. The local search power of the Fig. 6. Flowchart of Hill-climbing GA

9578
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

4. CASE STUDY
A CTSP with n  51 and m  4 is shown in Fig. 7, where V0
is the public city set (visit area) and V1-V4 are the private city
sets (visit areas) of Salesmen 1-4, respectively. All the
algorithms and processes are implemented in C++ on the
platform Microsoft Visual Studio 2010. The computer used is
Dell Inspiron620s having Windows 7 (32 bits) with CPU
Intel Corei3 and 2GB RAM at 3.30GHz.

Fig. 9. Evolution performance of GA with 20000 epochs

The total path length of the best individual of the initial
population of GA is about 1200km. With Greedy Algorithm,
the quality of the initial population can be greatly improved.
For example, Greedy GA and Hill-climbing GA converge
rapidly far before their preset termination condition. However,
it seems that the basic GA cannot complete its evolution
within 2000 epochs. Thus, we modify the generation count of
GA to be 20000. The convergence is extremely slow and the
evolution ends at about the 14000th epoch and the evolution
plot is shown in Fig. 9. Opposite to it, Greedy GA and Hill-
Fig. 7 CTSP and its city distribution climbing GA can accomplish their evolution at about the
Next, the three algorithms GA, greedy GA, and Hill-climbing 1300th generation. This is magnitude-fold saving in
GA are applied to solve the problem and their performances computational time.
are compared. We set the same parameters, i.e., the In addition, Hill-climbing GA outperforms Greedy GA.
individual number of a population to be 30, crossover Without a hill-climbing operation, the latter needs about 1200
probability 0.7, and mutation probability 0.1. Each algorithm epochs to evolve to the result with total tour length of
is run for five times and the maximum of epochs is 2000. 550(km); while the former spends about 200 epochs only to
achieve the same or better result.
4.1 Convergence Rate
4.2 Solution Quality
Experiments show that all the algorithms are convergent. To
compare their convergence rates, we plot the best individual The results obtained in the tests are listed in Table 1.
of each generation obtained by them as shown in Fig. 8.
Table 1. Results of three GAs (km)
Times GA Greedy GA Hill-climbing GA
1 641.867 547.884 526.461
2 672.578 550.629 522.525
3 662.864 549.26 529.311
4 650.236 550.637 519.256
5 616.612 556.521 520.921
Mean tour length 648.831 550.986 523.695

In TABLE 1, with the same set of parameters, spending 2000

epochs, Hill-climbing GA can reach the mean tour length of
523.695(km), compared to 648.831(km) of GA and
550.986(km) of Greedy GA.
From Fig.9, we find that without the Hill-climbing operation,
Fig. 8. Evolution performance of GAs with 2000 epochs GA traps in the local optimum and it is hard to reach its best

9579
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

result at about the 14000th epoch with the total tour length of REFERENCES
about 560(km). With the help of the Hill-climbing operation,
Basu, A., Elnaga, A., and Al-Hajj, A. (2000). Efficient
it is clear that after reaching the same result of Greedy GA,
coordinated motion. Mathematical and Computer
Hill-climbing GA can continue to obtain a better solution as
Modelling, vol. 31, pp. 39-53.
shown in Fig. 8.
Bektas, T. (Jun 2006). The multiple traveling salesman
The total tour length of the best solution with Hill-climbing problem: an overview of formulations and solution
GA is 519.256(km). The solution of visit tours is: procedures. Omega, vol. 34, pp. 209-219.
Carter, A.E., and Ragsdale, C.T. (2006). A new approach to
Salesman 1: 0→35→39→3→49→45→6→5→7→2→1→ solving the multiple traveling salesperson problem using
4→50→0; Salesman 2: 0→40→31→47→42→48→13→34 genetic algorithms. European journal of operational
→9→11→14→8→12→10→38→41→0; Salesman 3: 0→32 research, vol. 175, pp. 246-257.
→36→16→20→17→22→15→21→18→19→43→33→0; Cheong, T., and White, C.C. (2012). Dynamic traveling
and Salesman 4: 0→37→44→28→24→30→26→25→27→ salesman problem: Value of real-time traffic information.
23→29→46→0. IEEE Transactions on Intelligent Transportation Systems,
vol. 13, pp. 619-630.
The visit routes are shown in Fig. 10. Gutin, G., and Punnen, A. (2002). The Traveling Salesman
Problem and Its Variations. pp. 625-629. Kluwer,
Dordrecht.
Hirogaki, T., Aoyama, E., Ogawa, K., Hashimoto, N., and
Matsumura, M. (2005). CAM systems based on traveling
salesman problem from time perspective for high density
through-hole drilling. In Pts A-C (ed.), Advances in
Electronic Packaging, pp. 1783-1789.
Johnson, O., and Liu, J. (2006). A traveling salesman
approach for predicting protein functions. Source Code
for Biology and Medicine, vol. 1, pp. 1-7.
Li, J., Sun, Q.R., Zhou, M.C., and Dai, X.Z. (Oct 13-16,
2013). A New Multiple Traveling Salesman Problem and
its Genetic Algorithm-based Solution. In Proc. 2013
IEEE Int. Conf. Systems, Man and Cybernetics, pp. 1-6.
Manchester, UK.
Lim, A., Rodrigues, B., and Zhang, X. (2006). A simulated
annealing and hill-climbing algorithm for the traveling
Fig. 10. Visit tours of four salesmen tournament problem. European Journal of Operational
Research, vol. 174, pp. 1459-1478.
In summary, (1) Hill-climbing GA overcomes the deficiency
Ray, S.S., Bandyopadhyay, S., and Pal, S.K. (2007). Gene
of local search of GA to some extent by keeping its global
ordering in partitive clustering using microarray
search capability. It possesses better optimization ability and
expressions. Journal of Biosciences, vol. 32, pp. 1019-
yields better results than Greedy GA; and (2). The
1025.
convergence rate of Hill-climbing GA is the highest among
Ryan, J.L., Bailey, T.G., Moore, J.T., and Carlton, W.B. (Dec
the three considered GAs. This attributes to the introduction
13-16, 1998). Reactive tabu search in unmanned aerial
of the hill-climbing operation in GA.
reconnaissance simulations. In Proc. 30th Simulation
Conf. Winter. Washington, DC, vol.1, pp. 873-880.
5. CONCLUSION
Saleh, H.A., and Chelouah, R. (2004). The design of the
In this paper, we formulate a new multiple traveling salesman global navigation satellite system surveying networks
problem, CTSP, where different salesmen have different using genetic algorithms. Engineering Applications of
private city sets and share a set of public cities. It is Artificial Intelligence, vol. 17, pp. 111-122.
significant for modelling the applications where multiple Sun, Q.R. (2011). The research and application of the colored
individuals’ workspaces are not the same but partially traveling salesman problem. Master Dissertation, School
overlap with each other. To overcome the shortcomings of of Automation, Southeast University.
the classic GA, we present two improved genetic algorithms Tang, L., Liu, J., Rong, A., and Yang, Z. (2000). A multiple
(GA), called greedy GA and hill-climbing GA, by combining traveling salesman problem model for hot rolling
the classic one with the greedy algorithm and the hill- scheduling in Shanghai Baoshan Iron & Steel Complex.
climbing algorithm to solve CTSP. Finally, the presented European Journal of Operational Research, vol. 124, pp.
GAs are used to solve an example CTSP and the result shows 267-282,.
that the hill-climbing GA enjoys the best performance in Toth, P., Vigo, D.(Eds.) (2002). The Vehicle Routing
terms of convergence rate and solution quality. In the future, Problem. SIAM Monographs on Discrete Mathematics
we intend to research other heuristics of CTSP and explore and Applications. SIAM, Philadelphia.
the related applications.

9580