Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Second Asia International Conference on Modelling & Simulation

A Bee Colony Optimization Algorithm for Traveling Salesman Problem

Li-Pei Wongi, Malcolm Yoke Hean Lowii, Chin Soon Chongiii


i,ii
School of Computer Engineering, Nanyang Technological University, Nanyang Avenue,
SINGAPORE 639798.
iii
Singapore Institute of Manufacturing Technology, 71 Nanyang Drive, SINGAPORE 638075.
i
wonglipei@pmail.ntu.edu.sg, iiyhlow@ntu.edu.sg, iiicschong@SIMTech.a-star.edu.sg

Abstract The round-trip tour indicates that the salesman will


visit every city only once and then return to the starting
A Bee Colony Optimization (BCO) algorithm for city.
Traveling Salesman Problem (TSP) is presented in this The TSP is therefore to determine a Hamiltonian
paper. The BCO model is constructed algorithmically tour with minimum cost. TSP is one of the discrete
based on the collective intelligence shown in bee optimization problems which is classified as NP-hard
foraging behaviour. Experimental results comparing [2]. Generally, TSP can be denoted via graph notations
the proposed BCO model with some existing as follows:
approaches on a set of benchmark problems are • Let G = (V, E) be a graph.
presented. • V is a set of n cities, V = {v1, …, vn}.
• E is a set of arcs or edges,
1. Introduction E = {( r , s ) : r , s ∈ V } .
E is normally associated with a distance (or cost)
Many interesting collective behaviours can be matrix, which is defined as D = (dr,s). If dr,s = ds,r, the
observed in animals such as ant, bee, fish, cockroach problem is a symmetric TSP (STSP). Otherwise, it
etc. An example will be the highly coordinated patterns becomes an asymmetric TSP (ATSP). In TSP, our
shown in their foraging and aggregation behaviours. objective is to determine the permutation π of set V that
These individuals usually perform their actions locally minimizes Equation (1):
with limited knowledge about the entire system. n −1
However, when their local actions are combined, they C (π ) = ∑ dπ (i ),π (i +1) + dπ ( n ),π (1) (1)
emerge to produce global effects. i =1
This paper describes our work in applying the Bee Various techniques have been used to solve TSP. In
Colony Optimization (BCO) model, which is based on general, these techniques are classified into two broad
nectar collection in bee colony, to Traveling Salesman categories: exact and approximation algorithms [2].
Problem (TSP). This research is inspired by [1], on Exact algorithms are methods which utilize
using a new honey bee algorithm for dynamic mathematical models whereas approximation
allocation of Internet servers. The paper starts with a algorithms make use of certain heuristics and iterative
discussion on TSP in Section 2. A discussion on BCO improvements as the problem solving process. Some
model is presented in Section 3. Section 4 discusses instances in the exact methods category are Branch and
the implementation details of the BCO algorithm for Bound, Lagrangian Relaxation, Integer Linear
TSP. Experiments and results are presented in Section Programming etc.
5. Finally, this paper ends with conclusions and future The approximation algorithms can be further
works. classified into two groups: constructive heuristics and
improvement heuristics. Instances in constructive
2. Traveling Salesman Problem (TSP) heuristics include Nearest Neighbourhood, Greedy
Heuristics, Insertion Heuristics [3], Christofides
Suppose a salesman is given a set of cities Heuristics [4] etc. Instances in improvement heuristic
associated with traveling distances (or costs) from any include k-opt [5], Lin-Kernighan Heuristics [6,7],
city to any other city. The salesman is required to make Simulated Annealing [8], Tabu Search [9],
a round-trip tour with minimum distances (or costs).

978-0-7695-3136-6/08 $25.00 © 2008 IEEE 818


DOI 10.1109/AMS.2008.27
Evolutionary Algorithms [10-12], Ant Colony moves which are observed from the dance. This set of
Optimization [13,14], Bee System [15] etc. moves, named as “preferred path” and denoted as θ,
TSP is being studied by researchers from various will then serve as guidance in its foraging process. θ
disciplines because it is easy to initiate, difficult to basically contains a complete tour that had been
solve and has numerous applications. It is used as test explored previously by its mate and it will direct the
cases in this project as it is intended to explore the bee towards the destination. It is one of the
possibility of the proposed BCO model in solving permutation π of set V as described in Section 1,
problems that are similar to TSP nature. The θ ∈ Vπ .
development of this model will serve as an explanatory During the foraging process, a bee will travel from
example for such applications. In the subsequent one city to another city until it reaches the destination.
section, a BCO model will be presented. In the bee model, a heuristic transition rule is
employed to aid the bee in its decision making on
3. The BCO Model which city to be visited next. This rule consists of two
factors: arc fitness and heuristic distance. The arc
This section first describes the natural foraging fitness is computed for all possible paths to cities that
model of a typical bee colony. Next, an overview of can be visited by a bee from a particular city at a
how the foraging model helps in constructing feasible particular time. A higher fitness value is assigned to
paths for TSP is included. Focus will be given to the the arc which is part of the preferred path. By doing
arc fitness mechanism in the state transition rule. At this, a bee tends to choose the next visiting city based
the end of this section, a bee model for TSP is on the preferred path. On the other hand, under the
presented. heuristic distance influence, a bee tends to choose the
next visiting city which is nearest to its current city.
3.1. Bee Colony An example of a bee trying to visit five cities under
the influence of arc fitness will be presented. Let’s
The foraging behaviour in a bee colony remains assume five fully connected cities, denoted by V = {1,
mysterious for many years until von Frisch translated 2, 3, 4, 5}, are to be visited. The bee is assumed to start
the language embedded in bee waggle dances [16]. In foraging from its hive, which is denoted as city H. The
his series of experiments, he also discovered that bees bee is equipped with a preferred path, θ = “3, 2, 1, 4, 5,
actually own colour vision and attracted by scent 3”, which is observed via a dance before leaving the
deposited by their hive mates. hive. The hive and the rest of the cities are equally
To examine how bees communicate via waggle distanced from each other. By having the equidistance
dance, suppose a bee has found a rich food source. assumption, a bee is biased towards visiting the first
Upon its return to the hive, it starts to dance in a figure- city in the θ. It also ensures that the heuristic distance
eight pattern: a straight waggle run followed by a turn has no effect in the decision making in this example.
to the right back to the starting point, and then another When a bee is in city i at time t, two sets of next
straight waggle run followed by a turn to the left and visiting cities can be derived, namely the set of
back to the starting point again. The bee usually “allowed next cities”, Ai (t) and the set of “favoured
repeats these for a few times. Remarkably, via the next city”, Fi (t). Ai (t) contains all the feasible next
dance, the bee has actually informed its hive mates cities that can be reached from city i. Fi (t) contains a
about the direction and distance of the food source. single city which is favoured to reach from city i as
The direction is expressed via the angle of dance recommended by θ. Let θ(m) denotes the m-th element
relative to the sun position whereas the distance is in θ. If a bee has just started its exploration from the
expressed through the length of the straight waggle hive, FH(0) = {θ(1)}. If the current visiting city = θ(m)
run. This coding and decoding process will eventually at time t, then Fθ(m)(t) = {θ(m+1)}.
bring more bees towards the new food discovery. Some When t = 0, a bee leaves its hive and decides the
discussions about waggle dance can be found in first city to visit. Based on θ = “3, 2, 1, 4, 5, 3”, two
[17,18]. sets of cities are identified: AH(0) = {1, 2, 3, 4, 5} and
FH(0) = {3}. All the arcs that connect the hive and the
3.2. Path Construction in Artificial Bee cities in AH(0) will be assigned a fitness value. The
fitness assignment mechanism works as follows: λ is
In our proposed model, a bee is allowed to explore assigned to arc (H, 3) and (1- λ) / 4 is assigned to arcs
and search for a complete tour path. Before leaving the (H, 1), (H, 2), (H, 4) and (H, 5). Generally, the state of
hive, the bee will randomly observe dances performed fitness values at time t from city i can be summarized
by other bees. The bee is then equipped with a set of in an arc fitness matrix, Φi(t) with dimension 1 x |Ai (t)|.

819
Each entry in the matrix is denoted by ρij, which Bees are able to memorize how many bees visited
indicates the fitness value assigned to the arc from city certain links in the past and are able to advertise partial
i to city j where j ∈ Ai (t ) . The following matrix tours to other hive mates. Upon initiating a new
represents arc fitness values from hive to iteration, the solution is improved by 2-opt, 3-opt and
modified 3-opt heuristics.
city j ∈ A H ( 0 ) when t = 0:
In this paper, a variation of bee model for TSP will
Φ H (0) = [ ρ H 1 ρH 2 ρH 3 ρH 4 ρH 5] be presented as shown in Fig. 1.

1 − λ 1− λ 1− λ 1− λ  procedure BCO
= λ
 4 4 4 4  Initialize_Population( )
while stop criteria are not fulfilled do
λ represents the probability of following a city in while all bees have not built a complete path do
the preferred path. Notice that the arc fitness matrix is Observe_Dance( )
associated with the following property as shown in Forage_ByTransRule( )
Equation (2). Perform_Waggle_Dance ( )
∑ρ
j ∈ Ai ( t )
ij (t ) = 1 (2)
end while
end while
end procedure BCO
When t = 1 and let’s assume city 3 is chosen in the Figure 1. BCO model for TSP
previous stage: A3(1) = {1, 2, 4, 5} and F3(1) = {2}.
Likewise, the arc fitness of all cities in A3(1) is shown Several distinctions are identified between our
as below: proposed model and [15]:
Φ 3 (1) = [ ρ 31 ρ32 ρ34 ρ35 ] • Bees in our proposed model do not have the
ability to remember the number of bees that
1 − λ 1− λ 1− λ  have visited an arc.
= λ
 3 3 3  • Bees in our proposed model show the entire
The idea of the mechanism is to ensure that the city feasible path rather than partial tours via the
suggested by the preferred path is having higher waggle dance.
likelihood (λ in this case) to be chosen while the rest of • The bee hive in our proposed system is
the arcs are sharing the likelihood of 1 - λ. assumed to have equal distance to all cities.
Let’s assume when t = 1, city 5 is chosen. When • Bees are influenced by both arc fitness and
proceeding to t = 2, A5(2) = {1, 2, 4} and F5(2) = {3}. distance between cities when constructing
Notice that city 3 suggested by the preferred path is no solutions.
longer feasible as it has been visited at t = 0. Therefore, • Local search heuristics are not used to further
all the arcs that connect city 5 with the cities in A5(2) improve solutions found by bees.
will share equal likelihood of 1/3 as shown in the In our proposed model, a group of bees is created
following arc fitness matrix: during the initial stage. The number of bee, NBee, is
Φ 5 (2) = [ ρ 51 ρ52 ρ54 ] equal to the total number of cities in the TSP problem

[3 ]
instance. Subsequently, foraging process is initiated.
= 1 1 1 Bees are allowed to explore and exploit the search
3 3 space to construct a feasible path for TSP. When all the
bees have completed building a path, it is considered a
3.3. BCO Model for TSP complete bee cycle. As bees have no dance to follow
during the first bee cycle, they are allowed to explore
Various bee models have been attempted in the search space in random. This will allow
different domains ranging from dynamic server diversification in the search. The initialization phase
allocation for internet hosting center [1], numerical can be modified to restrict bees to construct paths
function optimization [19], hex game playing program based on certain constructive heuristics listed in
[20], job shop scheduling [21,22] to TSP [15]. In [15], Section 2.
Lucic and Teodorovic proposed a Bee System in As mentioned in Section 3.2, a bee is aided by a
tackling TSP. In their model, bees are expected to state transition rule in making its decision to choose
construct partial tours in every iteration. The position the next visiting city. The state transition probability,
of hive will be reallocated to a new city after every Pij(t), measures the likelihood to move from city i to
iteration. Interaction among bees in their model is city j at time t. It is a function of the city distance and
implemented in two ways: memory and waggle dance.

820
of the arc fitness present on the connecting edge. defined to be inversely proportional to the tour length.
Formally, it is defined in Equation (3): Bee with high profitability will dance for a longer
1 β duration.
[ ρ ij (t )]α ⋅ [ ]
d ij
Pij (t ) = (3) Table 1. Lookup Table for adjustment of Pfollow
1 Profitability Scores Pfollow

j∈Ai (t )
[ ρij (t )]α ⋅ [ ]β
d ij Pfi < 0.5Pfcolony 0.60
0.5Pfcolony ≤ Pfi < 0.65Pfcolony 0.20
ρij(t) is the arc fitness from city i to city j at time t and 0.65Pfcolony ≤ Pfi < 0.85Pfcolony 0.02
dij represents the distance between city i and city j. 0.85Pfcolony ≤ Pfi 0.00
Note that the city distance is inversely proportional to
Pij(t). In other words, the shorter the distance, the At the observing bee side, before it leaves the hive,
higher is the likelihood of that city to be selected. α is a it has to determine if it will follow the dance shown by
binary variable that turns on or off the arc fitness previous dancer with a probability of Pfollow. The
influence in the model. β is the parameter that controls probability Pfollow is adjusted dynamically according to
the significant level of heuristic distance. the profitability score of the bee and the colony based
Arc fitness, ρij(t), is defined in Equation (4): on the lookup table as shown in Table 1, which is
 λ , j∈Fi (t) adopted from [1,22]. In the extreme case where Pfollow
 1−λ⋅ A(t) ∩F(t) 
ρij(t) = ∀j∈Ai (t), 0≤λ≤1 is zero, the bee will keep to its own path.
, j∉Fi (t) (4)
i i
 A(t) − A(t)∩F(t)  Essentially, a bee is more likely to randomly
 i i i  observe and follow a waggle dance if its profitability
The mechanism of arc fitness in path construction rating is low compared to the average profitability of
has been discussed in Section 3.2. Notice that |Ai (t) ∩ the colony. Although a bee tends to be influenced
Fi (t)| will hold the value of 1 when there is a common either by its own or other bee’s experience, it may still
instance in both Ai(t) and Fi (t), or 0 otherwise. wander off from its preferred path occasionally.
Another important module in BCO is the waggle
dance. After a bee has built a complete path, a dance 4. Implementation Details
will be performed to other hive mates. The policy that
is applied in permitting waggle dances by returning The project is developed using JAVA with
bees is as follow: bees that construct shorter path NetBeans IDE 5.5 as the development tool. A list of
compared to its previous trials are allowed to dance. In tours is created to symbolize the dances by bees. Each
other words, not all bees are allowed to dance upon waggle dance tour is associated with a duration which
returning to hive. A bee has to make sure that it finds a is measured in terms of bee cycles. The duration will
better food source compared to its previous best trials determine how long the dance will be kept in the list.
in order to dance. If a bee dances, it will dance for The list will be refreshed and updated after every bee
certain duration. The dance duration of bee i, Di , is cycle to discard expired dances.
defined as in Equation (5): Our implementation utilizes the combination of
Pf i random exploration and the Nearest Neighbourhood
Di = K ⋅ (5) heuristic to generate the initial solutions during the first
Pf colony
bee cycle. Random exploration is implemented in such
1 a way that a bee starts from a random city, and chooses
Pf i = , Li = tour length (6)
Li the next visiting city by using the state transition rule,
which is discussed in Section 3.3. Nearest
1 n 1 n 1
Pf colony = ∑
n i =1
Pf i = ∑
n i =1 Li
(7) Neighbourhood heuristic suggests that a bee starts
from an arbitrary city and then chooses the nearest city
where K denotes the waggle dance scaling factor, Pfi to move ahead. Random selection is applied if there is
denotes the profitability score of bee i as defined in a tie in distance values observed in any step until a
Equation (6), Pfcolony denotes the bee colony’s average complete path is built. In the first bee cycle, a bee has
profitability as defined in Equation (7) and n denotes equal likelihood to employ one of the two approaches.
the number of bees performing waggle dance.
Pfi can be interpreted as the nectar quantity by bee i, 5. Experiments and Results
where bees are assumed to collect as much nectar as
possible. Higher quantity of nectar will be collected if This section first describes the benchmark problems
a bee travels along the shorter route. Therefore, Pfi is used in our experiments. It is followed by a discussion

821
on the benchmark algorithms which are included in our
comparison study. Finally, some preliminary Table 2. Performance of BCO Compared to ACS,
experimental results are presented. MMAS, GA, HGA and Lin-Ker
Instance EIL51 KROA100 LIN318
5.1. Benchmark Problems (Optimal) (426) (21282) (42029)
ACS [13] 0.00/0.48 0.00/0.65 n/a
MMAS [14] 0.00/0.16 0.00/0.10 n/a
The performance of the BCO model is investigated
GA [10,11] 0.00/0.00 0.00/0.00 0.00/0.00
by applying the algorithm to some of the benchmark HGA [12] 0.47/0.59 0.01/0.01 0.73/1.37
problems taken from TSPLIB1. 15 problem instances Lin-Ker [7,23] 0.00/0.00 0.00/0.00 0.00/0.08
are chosen from ATT, EIL, KRO and LIN series. ATT BS (2-opt) [15] 0.53/1.14 0.73/1.36 n/a
utilizes “pseudo-Euclidean” distance function as its arc BS (Modified 3- 0.00/0.00 0.00/0.00 n/a
weight type whereas EIL, KRO and LIN utilize opt) [15]
Euclidean distance function. The dimension of the BCO 0.47/0.85 2.26/3.43 6.32/7.55
problems ranges from 48 to 318 cities. x/y: x and y represents percentage difference of best tour length and
average tour length from optimal value respectively. “n/a” stands for
“not available”.
5.2. Benchmark Algorithms
Table 3. Performance of BCO on Benchmark
Six other approaches are included in our Problems Compared to Optimal Values Published
comparison study. They are: Ant Colony System in TSPLIB
(ACS) [13], MAX-MIN Ant System (MMAS) [14], Problem Optimal BCO
Genetic Algorithms (GA) [10,11], Hybrid of GA with Instances (% from optimal)
distance preserving crossover (HGA) [12], Lin- Best Average
Kernighan heuristic (Lin-Ker) [7,23] and Bee System ATT48 10628 0.31 0.83
(BS) [15]. The results reported in these papers are EIL76 538 0.19 2.01
included in Table 2 which will be discussed in Section EIL101 629 0.95 2.29
5.3. Notice that three different types of bee system KROB100 22141 2.24 3.10
were reported in [15]: bee system with 2-opt, 3-opt and KROC100 20749 0.50 1.50
modified 3-opt. However, only bee system with 2-opt KROD100 21294 1.64 3.25
and modified 3-opt are included in the table as the KROE100 22068 1.73 2.20
KROA150 26524 5.03 6.39
results for 3-opt and modified 3-opt are identical.
KROB150 26130 1.55 3.68
KROA200 29368 2.02 4.26
5.3. Results KROB200 29437 3.10 6.36
LIN105 14379 0.32 1.24
This section shows some preliminary results based
on the model which is discussed in Section 3. The The results show that BCO is able to achieve good
BCO model proposed in this paper is a stochastic performance for the three benchmark problems. While
algorithm. Hence, the results reported in this study are the results reported in both tables show that there is
the averages of five replications. The parameter still room for improvement in BCO model compared to
settings throughout the experiments are: BCMax = other approaches, we note that most of these
50000, NBee = total number of cities, α = 1, β = 10, λ = approaches make use of local search techniques to
0.99, K = 0.1 x BCMax. achieve optimality. These techniques will be
Table 2 and Table 3 summarize the experiment incorporated into the BCO in our future work to
results in this study. Table 2 highlights the comparison improve its performance.
of BCO model with other approaches. Three specific Table 3 shows that the performance of BCO
benchmark problems are employed for the comparison generally drops when size of the problem instances
study: EIL51, KROA100 and LIN318. The dimensions increase. There are a few reasons to this. One of them
of these problems are 51, 100 and 318 cities is the parameter settings which need to be fine-tuned to
respectively. Table 3 shows the results of BCO model cater for different scenarios. Our initial findings show
on the rest of the benchmark problems in the ATT, that the proposed BCO model spends most of the time
EIL, KRO and LIN series. exploring rather than exploiting good solutions. This
suggests that a more balanced strategy in exploring and
1
TSPLIB is a library of sample instances for the TSP and related exploiting the search space is needed.
problems. It is maintained by Reinelt, G. at Institut für Informatik,
Universität Heidelberg, Germany. URL: http://www.informatik.uni-
heidelberg.de/groups/comopt/software/TSPLIB95/

822
6. Conclusion [9]. J. Knox, "Tabu search performance on the symmetric traveling
salesman problem," Computers & Operations Research, vol.
21, no. 8, pp. 867-876, 1994.
A BCO model based on the bees’ collective [10]. B. Freisleben and P. Merz, "A genetic local search algorithm
foraging behaviour has been introduced to solve TSP. for solving symmetric and asymmetric traveling salesman
The model has been tested on a set of benchmark problems," in Proceedings of Internotional Conference on
Evolutionary Computation, 1996. pp. 616-621.
problems. As this is the initial implementation of BCO [11]. P. Merz and B. Freisleben, "Genetic local search for the TSP:
on TSP, we hope to improve the model further to New results," in Proceedings of the 1997 IEEE International
achieve optimal values for the list of benchmark Conference on Evolutionary Computation, 1997. pp. 159-164.
[12]. C. M. White and G. G. Yen, "A hybrid evolutionary algorithm
problems. Among the future enhancements are
for traveling salesman problem," in Proceedings of Congress
incorporating neighbourhood search in the model. on Evolutionary Computation, 2004, CEC2004., 2004. pp.
Neighbourhood search is useful when exploitation is 1473-1478.
desired. It can be applied after every bee cycle to [13]. L. M. Gambardella and M. Dorigo, "Solving symmetric and
asymmetric TSPs by ant colonies," in Proceedings of IEEE
enhance the quality of solutions. Constructive International Conference on Evolutionary Computation, 1996.
heuristics can also be integrated in preparing initial 1996. pp. 622-627.
solutions. This helps the BCO algorithm to have a [14]. T. Stützle and H. Hoos, "MAX-MIN ant system and local
better starting point in its search for optimal solutions. search for the traveling salesman problem," in Proceedings of
More bees’ related features such as the direction shown ICEC'97 - 1997 IEEE 4th International Conference on
Evolutionary Computation, 1997. pp. 308-313.
in waggle dance and scent sensing will be included to [15]. P. Lucic and D. Teodorovic, "Computing with Bees: Attacking
make the model more comprehensive. Another area in Complex Transportation Engineering Problems," International
our proposed model that will be investigated further is Journal on Artificial Intelligence Tools, vol. 12, no. 3, pp. 375-
394, 2003.
the parameter settings. Lastly, a study on improving
[16]. K. von Frisch, "Decoding the language of the bee," Science,
the model execution speed will be conducted by vol. 185, no. 4152, pp. 663-668, 1974.
making use of parallel and distributed processing. [17]. F. C. Dyer, "The biology of the dance language," Annual
Review of Entomology, vol. 47, pp. 917-949, 2002.
[18]. J. C. Biesmeijer and T. D. Seeley, "The use of waggle dance
Acknowledgement information by honey bees throughout their foraging careers,"
Behavioral Ecology and Sociobiology, vol. 59, no. 1, pp. 133-
We would like to thank Universiti Sains Malaysia 142, 2005.
for the scholarship awarded to Li-Pei Wong to pursue [19]. D. Karaboga and B. Basturk, "A powerful and efficient
his Ph.D. in Nanyang Technological University, algorithm for numerical function optimization: artificial bee
colony (ABC) algorithm," Journal of Global Optimization, vol.
Singapore. 39, no. 3, pp. 459-471, 2007.
[20]. J. V. Rijswijck, "Are bees better than fruitflies? Experiments
References with a hex playing program." in Proceeding of 13th Biennial
Conference of the Canadian Society for Computational Studies
[1]. S. Nakrani and C. Tovey, "On honey bees and dynamic server of Intelligence, AI 2000, Montreal, Quebec, Canada, May
allocation in Internet hosting centers," Adaptive Behavior, vol. 2000. Advances in Artificial Intelligence, vol. 1822/2000,
12, no. 3-4, pp. 223-240, 2004. Berlin / Heidelberg: Springer, 2007. pp. 13-25.
[2]. G. Laporte, "The traveling salesman problem: An overview of [21]. C. S. Chong, Y. H. Malcolm Low, A. I. Sivakumar, and K. L.
exact and approximate algorithms," European Journal of Gay, "A bee colony optimization algorithm to job shop
Operational Research, vol. 59, no. 2, pp. 231-247, 1992. scheduling," in Proceedings of the 2006 Winter Simulation
[3]. D. J. Rosenkrantz, R. E. Stearns, and P. M. Lewis, "An analysis Conference, 2006. pp. 1954-1961.
of several heuristics for the traveling salesman problem," SIAM [22]. C. S. Chong, Y. H. Malcolm Low, A. I. Sivakumar, and K. L.
Journal on Computing, vol. 6, no. 3, pp. 563-581, 1977. Gay, "Using a bee colony Algorithm for neighborhood search
[4]. A. M. Frieze, "An extension of Christofides heuristic to the k- in job shop scheduling problems," in Proceeding of 21st
person travelling salesman problem," Discrete Applied European Conference on Modeling and Simulation (ECMS
Mathematics, vol. 6, no. 1, pp. 79-83, 1983. 2007), 2007.
[5]. B. Chandra, H. Karloff, and C. Tovey, "New results on the old [23]. K. Helsgaun, "An effective implementation of the Lin-
k-opt algorithm for the traveling salesman problem," SIAM Kernighan traveling salesman heuristic," Roskilde University,
Journal on Computing, vol. 28, no. 6, pp. 1998-2029, 1999. 81, 1998.
[6]. S. Lin and B. W. Kerninghan, "An effective heuristic algorithm
for the traveling salesman problem," Operations Research, vol.
21, no. 2, pp. 498-516, 1973.
[7]. K. Helsgaun, "An effective implementation of the Lin-
Kernighan traveling salesman heuristic," European Journal of
Operational Research, vol. 126, no. 1, pp. 106-130, 2000.
[8]. E. H. L. Aarts, J. H. M. Korst, and P. J. M. Vanlaarhoven, "A
quantitative analysis of the simulated annealing algorithm - A
case study for the traveling salesman problem," Journal of
Statistical Physics, vol. 50, no. 1-2, pp. 187-206, 1988.

823

You might also like