Hindawi

Complexity
Volume 2017, Article ID 3813912, 12 pages
https://doi.org/10.1155/2017/3813912

Research Article
The Multiagent Planning Problem

Tamás Kalmár-Nagy,1 Giovanni Giardini,2 and Bendegúz Dezsy Bak1

1
Department of Fluid Mechanics, Faculty of Mechanical Engineering, Budapest University of Technology and Economics,
Budapest, Hungary
2
SPS Italiana Pack Systems, Novara, Italy

Correspondence should be addressed to Tamás Kalmár-Nagy; [email protected]

Received 31 July 2016; Revised 17 December 2016; Accepted 4 January 2017; Published 5 February 2017

Academic Editor: Roberto Natella

Copyright © 2017 Tamás Kalmár-Nagy et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.

The classical Multiple Traveling Salesmen Problem is a well-studied optimization problem. Given a set of 𝑛 goals/targets and 𝑚
agents, the objective is to find 𝑚 round trips, such that each target is visited only once and by only one agent, and the total distance
of these round trips is minimal. In this paper we describe the Multiagent Planning Problem, a variant of the classical Multiple
Traveling Salesmen Problem: given a set of 𝑛 goals/targets and a team of 𝑚 agents, 𝑚 subtours (simple paths) are sought such that
each target is visited only once and by only one agent. We optimize for minimum time rather than minimum total distance; therefore
the objective is to find the Team Plan in which the longest subtour is as short as possible (a min–max problem). We propose an
easy to implement Genetic Algorithm Inspired Descent (GAID) method which evolves a set of subtours using genetic operators.
We benchmarked GAID against other evolutionary algorithms and heuristics. GAID outperformed the Ant Colony Optimization
and the Modified Genetic Algorithm. Even though the heuristics specifically developed for Multiple Traveling Salesmen Problem
(e.g., 𝑘-split, bisection) outperformed GAID, these methods cannot solve the Multiagent Planning Problem. GAID proved to be
much better than an open-source Matlab Multiple Traveling Salesmen Problem solver.

1. Introduction from scheduling activities of companies and industries to

cooperative planning problems. See, for example, [17], where
Applications from space exploration [1–3] and drone delivery MTSP is used for modeling the preprinted insert scheduling
to search and rescue problems [4–7] have underlined the problem. Planning problems have also been investigated
need to plan a coordinated strategy for a team of vehicles to through MTSP formulations, specifically in [18, 19], where a
visit targets. It is important to develop a multiagent planner dynamic mission planning system for multiple mobile robots
for a team of autonomous vehicles to cooperatively explore operating in unstructured environments is presented (anal-
their environment [8, 9]. We formulate the overall planning ysis of planetary exploration), or in [20], where the MTSP
problem as finding a near-optimal set of paths that allows the formulation is used to describe a path planning problem for a
team of agents to visit a given number of targets in the shortest team of cooperative vehicles. An important and well-studied
amount of time. This problem is quite similar to the well- extension of the MTSP is the Vehicle Routing Problem [21,
known Multiple Traveling Salesmen Problem (MTSP) [10– 22], where a fleet of vehicles of different capacities, based at
13], a generalization of the Traveling Salesman Problem (TSP) either one or several depots, must deliver different customer
[14–16], that can be stated as follows: given 𝑛 nodes (targets) demands (the number of vehicles is often considered as a
and 𝑚 salesmen (agents), the MTSP consists of finding 𝑚 minimization criterion in addition to total traveled distance).
closed tours (paths start and end at the starting point of the A long-term goal of this work is to endow a team of
agents), such that each target is visited only once and by only autonomous agents (drones) with the capability of coopera-
one agent and the total cost of visiting all nodes is minimal. tive motion planning. In this application, the time available
MTSP has been used for modeling many real situations, for the solution is limited and real-time algorithms providing
2 Complexity

good approximate solutions are required. In this work, the with each edge is the Euclidean distance between the corre-
problem of planning a set of strategies for cooperatively sponding vertices (locations); that is, 𝑤(V𝑖 , V𝑗 ) = 𝑤(V𝑗 , V𝑖 ) =
exploring the environment with a fleet of vehicles is modeled ‖r(V𝑖 ) − r(V𝑗 )‖. The total cost 𝑐(⋅) of a path 𝑃 ∈ P𝑘 (𝐺) is the
as a variant of the classical MTSP, referred to as the Multiagent sum of its edge weights
Planning Problem (MAPP): given 𝑛 nodes (targets) and 𝑚
salesmen (agents) located at different depots, the MAPP seeks 𝑘

𝑚 tours such that each target is visited only once and by only 𝑐 (𝑃) = ∑𝑤 (𝑥𝑖 , 𝑥𝑖+1 ) . (4)
𝑖=1
one agent that minimizes a given cost function (specified by
(8) and (9)). The paper presents a Genetic Algorithm Inspired Analogously, for a cycle 𝐶 ∈ C𝑘 (𝐺)
Descent (GAID) method for obtaining good quality MAPP
solutions. 𝑘−1
The paper is organized as follows: after describing the 𝑐 (𝐶) = ∑ 𝑤 (𝑥𝑖 , 𝑥𝑖+1 ) + 𝑤 (𝑥𝑘 , 𝑥1 ) . (5)
MAPP in detail (Section 2) and an overview of how similar 𝑖=1
problems (MTSP in particular) are solved (Section 3), the
proposed GA-Inspired Descent method is described in Sec- After having introduced the necessary notation, we are
tion 4. Results are reported in Section 5, and conclusions are now in the position to formalize the combinatorial problems
drawn in Section 6. of interest. The Subtour Problem [25] is defined as finding a
simple path 𝑃 ∈ P𝑘 (𝐾𝑛+1 (𝑉)) of length 𝑘, starting at vertex
𝑥1 = 𝑎 and having the lowest cost 𝑐(𝑃) = ∑𝑘𝑖=1 𝑤(𝑥𝑖 , 𝑥𝑖+1 ).
2. The Multiagent Planning If 𝑘 = 𝑛, the problem is equivalent to finding the “cheapest”
Problem: Notations Hamiltonian path, where all the 𝑛 targets in 𝑇 are to be visited.
The general Traveling Salesman Problem, or 𝑘-TSP, poses
Graph theory [23, 24] provides a natural framework to
to find a simple cycle 𝐶 ∈ C𝑘+1 (𝐾𝑛+1 (𝑉)) of minimal cost
describe the Multiagent Planning Problem. Given 𝑉 =
starting and ending at vertex 𝑎, visiting 𝑘 targets.
{V1 , . . . , V𝑚 }, a set of 𝑚 elements referred to as vertices
Let 𝑇 = {𝑡1 , . . . , 𝑡𝑛 } be the set of 𝑛 targets (goals) to
(targets), and 𝐸 = {(V𝑖 , V𝑗 ) | V𝑖 , V𝑗 ∈ 𝑉}, a set of edges
be visited. The 𝑖th target 𝑡𝑖 is an object located in Euclidean
connecting vertices V𝑖 and V𝑗 , a graph 𝐺 is defined as the
space whose position is specified by the vector r(𝑡𝑖 ). Let 𝐴 =
pair (𝑉, 𝐸). Here we consider only undirected graphs, that is,
{𝑎1 , . . . , 𝑎𝑚 } denote the set of 𝑚 agents with position specified
graphs whose edges are unordered pairs with the symmetry
by r(𝑎𝑖 ). The classical Multiple Traveling Salesman Problem
relation (V𝑖 , V𝑗 ) = (V𝑗 , V𝑖 ). A complete graph is a graph where
can be formulated as follows. Let 𝑎 denote the unique depot;
all vertices of 𝑉 are connected to each other. The complete
that is, 𝑎𝑖 = 𝑎. The augmented vertex set is 𝑉 = 𝑇 ∪ 𝑎
graph induced by the vertex set 𝑉 is denoted by 𝐾𝑚 (𝑉), where
and the configuration space of the problem is the complete
𝑚 = |𝑉| (| ⋅ | denotes the cardinality of a set) is the number of
graph 𝐾𝑛+1 (𝑉). Let 𝐶𝑖 denote a cycle of length 𝑘𝑖 starting
vertices. A graph 𝐺1 = (𝑉1 , 𝐸1 ) is a subgraph of 𝐺 (𝐺1 ⊆ 𝐺) if
and ending at vertex 𝑎 (the depot). The Multiple Traveling
𝑉1 ⊆ 𝑉 and 𝐸1 ⊆ 𝐸 such that
Salesmen Problem is to find 𝑚 cycles (we also refer to these
𝐸1 = {(V𝑖 , V𝑗 ) | V𝑖 , V𝑗 ∈ 𝑉1 } . (1) as tours) 𝐶𝑖 of length 𝑘𝑖

A subgraph 𝑃 = (𝑉1 , 𝐸1 ) is called a path in 𝐺 = (𝑉, 𝐸) if 𝑉1 is C = {𝐶1 , . . . , 𝐶𝑚 } ,

a set of 𝑘 vertices of the original graph and 𝑚 (6)
∑𝑘𝑖 = 𝑛 + 𝑚,
𝐸1 = {(𝑥1 , 𝑥2 ) , (𝑥2 , 𝑥3 ) , . . . , (𝑥𝑘−1 , 𝑥𝑘 )} ⊆ 𝐸 (2) 𝑖=1

is the set of 𝑘 − 1 edges that connect those vertices. In other such that each target is visited only once (this also implies
words, a path is a sequence of edges with each consecutive visitation by only one agent) and the sum of the lengths
pair of edges having a vertex in common. Similarly, a (costs) of all the 𝑚 tours
subgraph 𝐶 = (𝑉2 , 𝐸2 ) of 𝐺 = (𝑉, 𝐸), with
𝑚
𝑉2 = {𝑥1 , . . . , 𝑥𝑘 } ⊆ 𝑉, 𝑐 (C) = ∑𝑐 (𝐶𝑖 ) (7)
(3) 𝑖=1
𝐸2 = {(𝑥1 , 𝑥2 ) , . . . , (𝑥𝑘−1 , 𝑥𝑘 ) , (𝑥𝑘 , 𝑥1 )} ⊆ 𝐸,
is minimal.
is called a cycle. The length of a path or cycle is the number of The Multiagent Planning Problem can be formulated
its edges. The set of all paths and cycles of length 𝑘 in 𝐺 will similarly. For the 𝑖th agent, the augmented vertex set is given
be denoted by P𝑘 (𝐺) and C𝑘 (𝐺), respectively. by 𝑉𝑖 = 𝑇 ∪ 𝑎𝑖 and the configuration space of the problem is
Paths and cycles with no repeated vertices are called the complete graph 𝐾𝑛+1 (𝑉𝑖 ). The weight associated with each
simple. A simple path/cycle that includes every vertex of the edge is the Euclidean distance between the corresponding
graph is known as a Hamiltonian path/cycle. Graph 𝐺 is locations, that is, 𝑤(V𝑖 , V𝑗 ) = 𝑤(V𝑗 , V𝑖 ) = ‖r(V𝑖 ) − r(V𝑗 )‖, with
called weighted if a weight (or cost) 𝑤(V𝑖 , V𝑗 ) is assigned to V𝑖 , V𝑗 ∈ 𝑉𝑖 , rendering 𝐾𝑛+1 (𝑉𝑖 ) a weighted and symmetric
every edge (V𝑖 , V𝑗 ). A weighted graph 𝐺 is called symmetric graph. Let 𝑃𝑖 denote a simple path (no repeated vertices) of
if 𝑤(V𝑖 , V𝑗 ) = 𝑤(V𝑗 , V𝑖 ). In this paper, the weight associated length |𝑃𝑖 | = 𝑘𝑖 starting at vertex 𝑎𝑖 . The optimal solution of
Complexity 3

the Multiagent Planning Problem (MAPP) is a set P of 𝑚 Genetic Algorithms are used to minimize the sum of the
pairwise disjoint (modulo the starting points) paths 𝑃𝑖 salesmen path lengths together with the maximum distance
traveled by each salesmen (to balance the agent workload).
P = {𝑃1 , . . . , 𝑃𝑚 } , A Modified Genetic Algorithm on the equivalent TSP is
𝑚 used in [32]. Shirafkan et al. [45] proposed a hybrid method
(8)
∑𝑘𝑖 = 𝑛, incorporating simulated annealing and Genetic Algorithm
𝑖=1 to solve the multidepot MTSP. In this MTSP variant the
starting location (depot) of each agent is fixed and multiple
such that the length of the longest path agents can have start from the same depot. The multiobjective
MTSP variant is presented in [46] and Genetic Algorithm is
𝑐max (P) = max 𝑐 (𝑃𝑖 ) (9) used to find approximate solutions. In multiobjective MTSP
𝑖
the total distance traveled by the agents and the balance of
is minimal (a min–max problem). In other words, the 𝑚 agent working times are the objectives. Kaliaperumal et al.
agents have to visit 𝑛 targets in the least amount of time, and [47] also used Genetic Algorithm to solve the MTSP. They
every target is visited only once. The number of targets visited developed a brand new crossover operator called modified
by each agent can be different, but each agent has to visit at two-part chromosome crossover and obtained better results
least one target (i.e., 𝑘𝑖 ≥ 2). using this new operator. A new (branch-and-cut type) exact
method is presented in [48] for the heterogeneous MTSP
3. Overview of Solution Methods (some targets can be visited only by a specific agent). In [49]
a heuristic search method is proposed to solve the Multiagent
The Multiple Traveling Salesmen Problem (MTSP) and the Path Finding problem, which is similar to MAPP, except that
Multiagent Planning Problem are notoriously difficult to the endpoints of the tours are also fixed. These methods
solve due to their combinatorial nature (they are NP-hard). work well on the problem they were constructed for, but
A common approach is to transform the MTSP into an none of them is designed to solve the MAPP. In Section 5.1.2
equivalent Traveling Salesman Problem, for which solutions we benchmark our GAID against some of these methods in
can be found by exact methods (e.g., branch-and-bound MTSP solving. In Section 5.2 we compare GAID against a
algorithms and linear programming [26–28]) or approximate Matlab solver which is capable of solving MAPP as well as
algorithms such as Genetic Algorithms, Simulated Anneal- MTSP.
ing, and Ant System [29, 30]. For example, in [31, 32] the
authors proposed to transform the MTSP into an equivalent
TSP by adding dummy cities and edges with ad hoc null or 4. The Multiagent Planning Problem:
infinite costs. However, as stated in [33–35], transforming the Approximate Solution
MTSP into an equivalent TSP might yield an even harder
problem to solve. Similar approaches are investigated in [13, Given 𝑚 ≥ 1 agents and 𝑛 known targets/cities to visit, the
36, 37]. optimal team strategy (called Team Plan) is sought that allows
The first attempt to solve large-scale MTSPs is given in the fleet to visit every target only once.
[33], where a branch-and-bound method (the most widely We represent the Team Plan as a collection of 𝑚 distinct
adopted technique for solving these combinatorial problems subtours. Thus, given 𝑚 agents and 𝑛 targets, Team Plan
[38]) is applied to both Euclidean (up to 100 cities and 10 P is defined as P = {𝑃1 , . . . , 𝑃𝑚 }, where 𝑃𝑖 is the path of
salesmen) and non-Euclidean problems (up to 500 cities and the 𝑖th agent visiting 𝑘𝑖 < 𝑛 targets. Note that this rep-
10 salesmen). Branch-and-bound is also applied in [39] for resentation allows the individual subtours to have different
solving an asymmetric MTSP up to 100 cities. lengths. Moreover, the Multiagent Planning Problem can also
Other solution methods have also been proposed, for be rewritten for each 𝑖th agent to find the best possible
example, simulated annealing [18], Gravitational Emulation subtour of length 𝑘𝑖 < 𝑛 that satisfies the imposed cost
Local Search (GELS) algorithm [40], and evolutionary algo- function. We propose an optimization technique we call the
rithms. In [41], different evolutionary algorithms, ranging Genetic Algorithm Inspired Descent (GAID) method. Briefly,
from Genetic Algorithms to Particle Swarm and Monte- a Genetic Algorithm (GA) is an optimization technique used
Carlo optimization, are compared. The MTSP with ability to find approximate solutions of optimization and search
constraint is solved with an Ant Colony Optimization (ACO) problems [50–52]. Genetic Algorithms are a particular class
algorithm in [34], where the MTSP is not translated into of evolutionary methods that use techniques inspired by
an equivalent TSP and the ACO algorithm is modified for Darwin’s theory of evolution and evolutionary biology, such
dealing with the characteristics of the original problem. In as inheritance, mutation, selection, and crossover. In these
[34] results are compared with a Modified Genetic Algorithm systems, populations of solutions compete and only the fittest
that solves the equivalent TSP. Linear Programming is used survive.
in [35], and similarly to [34], the original MTSP is analyzed Similarly to the classical GA, GAID consists of two
and solved. In both [34, 35], the authors conclude that phases: initialization and evolution. In the initialization
the original MTSP is easier to solve than the derived TSP. phase, the starting Team Plan is created (see Section 4.1),
An important work is [42], where different local search while the evolution phase (see Section 4.2) evolves Team Plan
heuristics are presented and compared. In [17, 29, 43, 44] P toward a better quality final solution.
4 Complexity

4.1. Initialization Phase. During the initialization phase, the application depends on a given probability, as shown in
starting Team Plan—a starting set of subtours—is created. Let Figure 1.
𝑇1 = {𝑡1 , . . . , 𝑡𝑛 } and 𝐴 = {𝑎1 , . . . , 𝑎𝑚 } be the sets of 𝑛 targets The crossover operator (Figure 2) is applied with prob-
and the 𝑚 agents, respectively. Without loss of generality, we ability 𝑝crossover and combines the genetic materials of two
assume that the order of planning is 𝑎1 , 𝑎2 , . . . , 𝑎𝑚 and that selected subtours (called parents), replacing them with the
the starting subtours have similar lengths. A Team Plan is two newly created ones (the offsprings). Parents are chosen
feasible if the subtours are pairwise disjoint (except possibly using the best-worst selection with probability 𝑝best-worst
their starting points). (Figure 2(a)) or randomly (Figure 2(b)) with probability 1 −
At first, a subtour 𝑃1 (with starting point 𝑎1 ) of length 𝑘1 is 𝑝best-worst . The parents are mated the classical way [50]: they
chosen. Then subtour 𝑃2 is chosen from the target set 𝑇2 that are randomly halved (not necessarily at the same position)
is simply obtained by discarding from 𝑇1 the targets visited and the halves are simply swapped.
by agent 𝑎1 . In general, the 𝑖th agent plans a subtour 𝑃𝑖 on the The Mutation Operator changes the Team Plan (with
targets not yet allocated previously. Obviously, this process probability 𝑝mutation ) by randomly swapping two genes
yields a feasible Team Plan P = {𝑃1 , . . . , 𝑃𝑚 }. between two different subtours. The Migration Operator
Here we utilize three simple initialization methods. (applied with probability 𝑝migration ) removes a randomly
chosen target from subtour 𝑃𝑖 (of length 𝑘𝑖 ) and adds it to
(i) Random initialization: for each agent the targets are subtour 𝑃𝑗 (of length 𝑘𝑗 ). The location at which this target
selected at random from the unallocated ones. is placed into subtour 𝑃𝑗 is also chosen at random. Note
(ii) Greedy initialization: the initial Team Plan is created that the lengths of the subtours change: 𝑘𝑖 decreases while 𝑘𝑗
using a greedy approach to form feasible starting increases.
subtours. Each agent selects its targets using a Nearest With probability 𝑝boost , each subtour is processed with
Neighbor heuristics: for a given a target, the next a heuristic boosting technique. The 2-opt method [53–55]
target will be the nearest one. is adopted here to directly improve the quality of subtours
(it replaces subtours with better ones from their “neighbor-
(iii) TSP-based initialization: for problems where the posi- hood”). The 2-opt method determines whether the inequality
tions of the 𝑚 agents are not imposed (the 𝑚 agents
can start from any target 𝑡𝑖 ∈ 𝑇), a feasible starting
𝑐 (𝑥𝑖 , 𝑥𝑖+1 ) + 𝑐 (𝑥𝑗 , 𝑥𝑗+1 ) > 𝑐 (𝑥𝑖 , 𝑥𝑗 ) + 𝑐 (𝑥𝑖+1 , 𝑥𝑗+1 ) (10)
Team Plan can be generated by clustering the TSP
solution computed on the complete graph 𝐾𝑛 (𝑇).
between the four vertices 𝑥𝑖 , 𝑥𝑖+1 , 𝑥𝑗 , and 𝑥𝑗+1 of a
“Clustering” is carried out by discarding 𝑚 edges
path/subtour holds, in which case edges (𝑥𝑖 , 𝑥𝑖+1 ) and
from the TSP tour, in order to subdivide it “fairly”,
(𝑥𝑗 , 𝑥𝑗+1 ) are replaced with edges (𝑥𝑖 , 𝑥𝑗 ) and (𝑥𝑖+1 , 𝑥𝑗+1 ),
and having 𝑚 starting subtours with similar costs.
respectively. This method provides a shorter path without
This initialization method introduces a degree of
intersecting edges.
complexity in the overall system, since a TSP solution
must be computed.
5. Results
The initialization phase is an important step in the
optimization process, since it directly influences the efficacy An extensive set of simulations were run to test the per-
of the algorithm. formance of the proposed GAID method. Unless otherwise
specified, simulations were run for 150000 generation steps,
4.2. Evolution Phase. The evolution phase evolves the Team while the crossover, mutation, migration, and boosting (2-
Plan, trying to design a strategy where the overall time is opt) operators were applied with probabilities 𝑝crossover = 0.7,
reduced (minimizing the cost 𝑐max (P); see (9)). 𝑝mutation = 0.4, 𝑝migration = 0.6, and 𝑝boost = 0.3, respectively.
This phase has the same mechanism of a classical Genetic For the application of the crossover operator, the probability
Algorithm [50] with one important difference: there is only of best-worst selection is 𝑝best-worst = 0.5 (thus the probability
one Team Plan P to be evolved (i.e., the population size is 1). for random selection is 1 − 𝑝best-worst = 0.5).
At every evolution/generation step, a set of operators (see
Section 4.3) is applied to the subtours 𝑃𝑖 ∈ P. If P improves, 5.1. Comparison with Structured and Known Solutions. In this
it is kept for the next generation step, otherwise the previous section, GAID is compared with known results. In order to
Team Plan is restored (we expect that a simulated annealing- make the comparison meaningful, we also imposed the same
type modification will be even more efficient). The flowchart constraints (if any) used in the referenced works.
of the GAID evolution phase is shown in Figure 1.
5.1.1. An Example Problem with Known Solution. In this
4.3. Team Plan Operators and Boosting. The evolution of P section we present a 9-81 MAPP (9 agents, 81 targets), in
toward a near-optimal multiagent strategy is accomplished by which the agents and targets form 9 tightly packed clusters
combining the genetic materials of its subtours through the as shown in Figure 3. Each cluster contains an agent and 9
application of genetic-like operators. Three different opera- targets. Since the distances among the clusters are consid-
tors have been designed: crossover, mutation, and migration. erably longer than those among the points of a cluster, the
The operators are applied in a predefined order, and their optimal solution can be found by treating each cluster as a
Complexity 5

Starting
Team Plan

Minimum cost
Team Plan

Subtours Genetic Team Plan Team Plan

selection operators evaluation comparison

Crossover Mutation Migration Local 2-OPT

operator operator operator boosting
Parent A Offspring A
Pcrossover Pmutation P migration Pboost
Parent B Offspring B

Figure 1: Flowchart of the evolution phase, together with the sequence of operators.

Team Plan Temporary Team Plan Team Plan Temporary Team Plan
Plan A → Best A H Plan A Plan A
Plan B Plan B Plan B Plan B
Plan C Plan C Plan C C G
Plan D Plan D Plan D Plan D
Crossover Crossover
Plan E Plan E Plan E Plan E
Plan F Plan F Plan F Plan F
Plan G Plan G Plan G G C
Plan H → Worst H A Plan H Plan H
(a) Best-worst selection (b) Random selection

Figure 2: Crossover operator. (a) In case of best-worst selection the best and the worst subtours are mated (with probability 𝑝best-worst ). (b)
By random selection the two subtours are selected at random (with probability 1 − 𝑝best-worst ).

1-9 MAPP (1 agent, 9 targets). The tour lengths are between 5.1.2. Comparison with Evolutionary Algorithms. We also
356.1 and 548.6, meaning that the optimal cost (the length compare our method with the results reported in [34], where
of the longest tour) is 548.6 (see (9)). In Figure 3 we show an Ant Colony Optimization algorithm is compared with the
the random and greedy initialization for this MAPP. In both Modified Genetic Algorithm (MGA) of [32]. We note here
cases some of the subtours contain a higher number of targets that these methods cannot solve the MAPP, therefore six
than others. We intentionally set the initialization this way to MTSPs, all taken from well-known TSPLIB [56], are solved.
further challenge the algorithm. The initial tour lengths vary The number of agents is set to 𝑚 = 5, and their starting
between 1944.9–15302.3 for the random, and 175.3–1627.9 for location is the first target in the corresponding problem
the greedy initialization. Consequently, the initial Team Plan data file. Since a single depot is considered, only the greedy
costs are 15302.3 and 1627.9, respectively. initialization method is used. In both [32, 34], rounded costs
The convergence of the Team Plan cost for both simu- are considered. Therefore, for meaningful comparison, we
lations is shown in Figure 4 together with the cost of the need to minimize the following cost function:
optimal solution (548.6). At the end of the simulations the
𝑚
costs of the subtours varied between 378.3–612.3 and 389.9–
𝑐round (C) = ∑round (𝑐 (𝐶𝑖 )) . (11)
594.1 for solutions starting from randomly initialized and 𝑖=1
greedy initialized Team Plans, respectively.
The final Team Plans of the two simulations are shown In addition, the maximum length of the cycles, 𝑘max , is
in Figure 5. We see that those long edges which initially con- limited to 20 (note that since MTSPs are considered here, we
nected targets belonging to different clusters were completely need to modify the Team Plan using cycles instead of paths).
eliminated. The longest subtour connects the targets in the Results based on 100 simulations are presented in Table 1. In
middle-top cluster for both simulations. this case, our results outperform those reported in the above
6 Complexity

Table 1: Comparison between the proposed GAID, the Ant Colony Optimization, and the Modified Genetic Algorithm methods. Results are
averaged over 100 simulations. The number of targets is included in the TSPLIB problem name. 𝑘max is the maximum number of targets an
agent can visit. For each case, the fixed number of 𝑚 = 5 agents is considered, starting from the same location (one-depot problem). Only
the greedy initialization method is used.

ACO MGA GAID Difference of means

TSPLIB Problem 𝑘max
Min Mean Min Mean Min Mean ACO MGA
pr76 20 178597 180690 157444 160574 152722 156503.9 −15.4% −2.6%
pr152 40 130953 136341 127839 133337 114698 126128.8 −8% −5.7%
pr226 50 167646 170877 166827 178501 152198 158073.9 −8% −12.9%
pr299 70 82106 83845 82176 85796 70059 71705.1 −16.9% −19.6%
pr439 100 161955 165035 173839 183698 136169 138655.5 −15.6% −24.5%
pr1002 220 382198 387205 427269 459179 311492 319240.4 −17.5% −30.5%

2000 2000

1500 1500

y y
1000 1000

500 500

500 1000 1500 2000 500 1000 1500 2000

x x
(a) 9-81 MAPP: initial plan with random initialization (b) 9-81 MAPP: initial plan with greedy initialization

Figure 3: The GAID algorithm was tested on a MAPP which contains 81 targets forming 9 clusters on a 2200 × 2200 square domain. An
agent was placed into each cluster (white diamonds); thus the number of agents is 𝑚 = 9. The initial cost based on (9) is 15302.3 for random
and 1627.9 for greedy initialization. The longest subtour is highlighted with thick line.

references. Figure 6 also shows the solution obtained for the For each test case, our results are compared with only the best
pr76 TSPLIB problem. ones of [42]. We also report the name of the Heuristic with
which each referenced solution has been obtained (please
5.1.3. Comparison with Heuristics. We compare our method refer to [42] for a more detailed description of the adopted
against the results reported in [42], where different heuristics heuristic methods). Tables 2 and 3 show the results obtained
for solving MTSP instances are proposed and compared. by initializing the Team Plan with either the greedy or the
In [42], a no-depot MTSP variant is considered (the TSP-based initialization methods, respectively. Results are
agents do not have a predefined starting location), and the averaged over 100 simulations.
min–max optimization problem is solved ((9) is minimized). In each test case, GAID returns good solutions, regardless
In addition, the number of salesmen is fixed and no con- of the applied initialization method. In general, our best
straints on the plan lengths are imposed. Note that since solutions are close to the ones from the literature and in
we are comparing our method with MTSP results, we need a few cases (e.g., in the berlin52 or the pr264 problems)
to modify our Team Plan accordingly, using cycles instead are even better. The greedy initialization method seems to
of paths. In addition, since in [42] rounded distances are provide a better initial Team Plan in those problems where the
considered, for meaningful comparison the cost function (see distribution of targets has a well-defined structure (the pr264
(9)) is modified accordingly: problem), while in “small-size” problems (i.e., berlin52)
the TSP-based method is preferable (even if the obtained
𝑐maxround (C) = max round (𝑐 (𝐶𝑖 )) . (12)
𝑖 improvement does not justify its required complexity, time,
Complexity 7

Table 2: Heuristics comparison for 100 simulations. The greedy initialization method was used.

Heuristics Greedy initialization Errors

TSPLIB Problem Number of Agents
Method Minimum Mean Minimum Mean Minimum Mean
4 Bisection 2182 2204.3 2183 2422.4 0.04% 9.8%
5 𝑘-split 1713 1739.7 1825 2160.4 6.5% 24.2%
berlin52
SGH 1531 1585 — 20.2%
6 1611 1905.8
Bisection 1476 1643.1 9.1% —
4 SGH 5955 6096.7 6000 6690.3 0.7% 9.7%
5 𝑘-split 4629 5025.9 4964 5620.1 7.2% 11.8%
kroA100
𝑘-center 4200 4429.4 4% —
6 4370 5038.4
𝑘-means 4230 4234.6 — 18.9%
4 Bisection 32423 32757.5 32434 35740.9 0.03% 9.1%
bier127
6 𝑘-split 22815 23071.7 24608 26993.3 7.8% 16.9%
4 Bisection 12705 12705 12196 13830.3 −4.1% 8.8%
pr264 𝑘-means 8526 9131.6 5.5% —
6 9000 10256.3
Bisection 8739 9051.6 — 13.3%

Table 3: Heuristics comparison for 100 simulations using the TSP-based initialization.

Heuristics TSP-based initialization Errors

TSPLIB Problem Number of Agents
Method Minimum Mean Minimum Mean Minimum Mean
4 Bisection 2182 2204.3 2088 2359.6 −4.5% 7%
5 𝑘-split 1713 1739.7 1804 2014.8 5.3% 15.8%
berlin52
SGH 1531 1585 — 13.6%
6 1576 1801.2
Bisection 1476 1643.1 6.7% —
4 SGH 5955 6096.7 6000 6443.6 0.7% 5.7%
5 𝑘-split 4629 5025.9 4796 5314.6 3.6% 5.7%
kroA100
𝑘-center 4200 4429.4 2.6% —
6 4310 4693.9
𝑘-means 4230 4234.6 — 10.8%
4 Bisection 32423 32757.5 32948 34606 1.6% 5.6%
bier127
6 𝑘-split 22815 23071.7 24290 26497.1 6.4% 14.8%
4 Bisection 12705 12705 13400 19135.3 5.4% 50.6%
pr264 𝑘-means 8526 9131.6 1.2% —
6 8629 14395.8
Bisection 8739 9051.6 — 59%

and computational costs). Even though these heuristics are

10 4 superior in solving MTSP compared to GAID, they cannot
solve MAPP, while our method can solve both.
5000 Figure 7 shows the solutions obtained by applying the
greedy (Figure 7(a)) and the TSP-based (Figure 7(b)) ini-
Cost

tialization methods for the kroA100 problem (with 𝑚 = 5

agents). Compared with [42], the TSP-based method and the
greedy method result in final solutions that are only 3.6% and
1000 7.2% worse than the cited one, respectively.

500
10 100 1000 104 105 5.2. Comparison with Other Software. In this section we test
Generation GAID against a freely available Matlab MTSP solver based on
a Genetic Algorithm [57].
Random initialization The number of agents is fixed, and the minimum path
Greedy initialization size, 𝑘min = 2, is imposed (this way, solutions with paths
Optimal cost
composed of only one target are avoided). The cost function,
Figure 4: Convergence of the Team Plan cost for the 9-81 MAPP which is specified by (9), is minimized, and targets are ran-
shown in Figure 3. domly generated over the unit square map. In all simulations,
8 Complexity

2000 2000

1500 1500

y y
1000 1000

500 500

500 1000 1500 2000 500 1000 1500 2000

x x
(a) 9-81 MAPP problem: final plan after random initialization (b) 9-81 MAPP problem: final plan after greedy initialization

Figure 5: The GAID algorithm found 9 subtours which connect targets belonging to the same cluster for the 9-81 MAPP with both
initialization strategies. The final costs are 612.3 and 594.1, respectively.

14000 14000
12000 12000
10000 10000
y coordinate

y coordinate

8000 8000
6000 6000
4000 4000
2000 2000
0 0
0 5000 10000 15000 20000 0 5000 10000 15000 20000
x coordinate x coordinate
(a) pr76 TSPLIB problem: starting Team Plan (b) pr76 TSPLIB problem: final Team Plan

Figure 6: The MTSP is solved for the pr76 TSPLIB target configuration. The maximum plan length is 𝑘max = 20 targets and the number of
agents is 𝑚 = 5. The cost function specified by (11) is minimized, and the comparison between the starting and the final Team Plans is shown.
In (a) the starting Team Plan, with cost equal to 194618, is shown, while in (b) the final Team Plan, with cost equal to 152722, is reported.
These solutions are 15.4% and 2.6% better than the referenced ones, obtained by implementing an Ant Colony Optimization method and a
Modified Genetic Algorithm, respectively.

the greedy initialization method is adopted. Since the Matlab such that each subtour visits 10 targets. The simulation was
MTSP solver is based only on a Genetic Algorithm and no ran for 1000000 generations.
heuristics is used, we also run a set of tests with 𝑝boost = 0%. Figures 9 and 10 summarize the results. The initial cost
Table 4 shows the obtained results, averaged over 100 of 1.94 was reduced by 43.5% (to 1.08) by the end of the
simulations. Our method clearly outperforms the Matlab 150000th generation and by a further 5.7% (to 0.94) till the
solver (clearly, with the 2-opt method the results are much end of the simulation.
better). Figure 8 shows two examples where the Matlab MTSP
solver solution is compared with the GAID ones.
6. Conclusions and Future Work
5.3. Solution for a Large Example Problem. We tested our This paper describes the Multiagent Planning Problem, a
algorithm on a 20-400 MAPP. The initial Team Plan was variant of the classical Multiple Traveling Salesman Problem.
constructed by means of the greedy initialization method, Our solution method uses a simplified Genetic Algorithm,
Complexity 9

2000 2000

y coordinate
y coordinate

1000 1000

0 0
0 2000 4000 0 2000 4000
x coordinate x coordinate
(a) kroA100 problem: greedy initialization (b) kroA100 problem: TSP-based initialization

Figure 7: Solving the MTSP by the greedy and the cluster initialization methods. In both cases, 𝑚 = 5 agents are considered, and the target
set is kroA100. In (a), the greedy initialization method is used and the final obtained Team Plan is shown. The cost is 4964, and it is 7.2%
worse than the referenced one. In (b), the TSP-based initialization method is used and the obtained final Team Plan is shown. The cost is
4796, which is only 3.6% worse than the referenced one.

1 1
y coordinate

y coordinate

0.5 0.5

0 0
0 0.5 1 0 0.5 1
x coordinate x coordinate
(a) Matlab solution (b) Proposed method

Figure 8: GAID has been compared with the Matlab MTSP solver. In this example, MAPP is solved for 100 targets randomly distributed in
the unit square. 𝑚 = 10 agents are considered. The minimum path length is 𝑘min = 2. The cost function (see (9)) is minimized. In (a) the
solution obtained by running the MTSP Matlab solver is shown, with cost 7.92. In (b) our solution is shown with cost 6.37, which is 19.5%
better than the Matlab MTSP solution.

called Genetic Algorithm Inspired Descent (GAID), capable here show the success of the approach, demonstrating how
of solving both the MTSP and the MAPP. GAID takes an a simple method can solve an otherwise hard combinatorial
initial Team Plan and applies various genetic operators to problem.
decrease the length of the longest path. We benchmarked
GAID against other evolutionary algorithms and heuristics. Competing Interests
GAID outperformed the Ant Colony Optimization and the
Modified Genetic Algorithm. Even though the heuristics The authors declare that they have no competing interests.
specifically developed for MTSP (e.g., 𝑘-split, bisection)
outperformed GAID, these methods cannot solve MAPP. Acknowledgments
GAID proved to be much better than the Matlab MAPP
solver. GAID is also easy to implement. This work relates to the scientific programs “Development
A long-term goal of this work is to endow a team of of quality-oriented and harmonized R+D+I strategy and the
autonomous agents (drones) with the capability of coopera- functional model at BME” (Project ID: TÁMOP-4.2.1/B-
tive motion planning. In this application, the time available 09/1/KMR-2010-0002) and “Talent care and cultivation in the
for the solution is limited and fast algorithms providing scientific workshops of BME” (Project ID: TÁMOP-4.2.2/B-
good approximate solutions prevail. The results presented 10/1-2010-0009).
10 Complexity

1.0 1.0

0.8 0.8

0.6 0.6
y y

0.4 0.4

0.2 0.2

0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0
x x
(a) 20-400 MAPP: initial Team Plan with greedy initialization (b) 20-400 MAPP: final Team Plan after the 1000000th generation

Figure 9: The GAID algorithm was performed on a MAPP with 20 agents and 400 targets located in the unit square. The Team Plan cost
was 1.91 after the initialization and was reduced to 1.08 by the end of the 150000th generation and to 0.94 by the end of the simulation. The
longest subtour is highlighted by a thick line.

Table 4: Comparison between the proposed GAID method and a Matlab MTSP code. For each test, 𝑘min = 2 and the greedy initialization
method is used. Targets are randomly generated over the unit square.

Matlab code Proposed method Cost variation

Test case
Min Mean Max 𝑝boost Min Mean Max Min Mean Max
100 targets 30% 6.37 6.72 7.11 −10.7% −17.9% −22.8%
7.14 8.19 9.21
10 agents 0% 6.67 7.14 7.58 −6.5% −12.8% −17.7%
200 targets 30% 9.55 9.89 10.25 −39% −43.5% −47.8%
15.64 17.5 19.65
20 agents 0% 10.05 10.6 11.23 −35.7% −39.4% −42.8%

2.0 [2] S. Hayati, R. Volpe, P. Backes et al., “The Rocky 7 rover: a Mars
sciencecraft prototype,” in Proceedings of the IEEE International
1.8 Conference on Robotics and Automation (ICRA ’97), vol. 4,
Albuquerque, NM, USA, April 1997.
1.6 [3] E. T. Baumgartner, “In-situ exploration of Mars using rover
Cost

systems,” in Proceedings of the AIAA Space 2000 Conference,

1.4 2000-5062, Long Beach, Calif, USA, September 2000.
[4] A. Birk and S. Carpin, “Rescue robotics—a crucial milestone on
1.2 the road to autonomous systems,” Advanced Robotics, vol. 20,
no. 5, pp. 595–605, 2006.
1.0 [5] S. Sarieland and H. L. Akin, “A novel search strategy for
1 10 100 1000 autonomous search and rescue robots,” in Robo Cup 2004: Robot
104 105 106
Soccer World Cup VIII, vol. 3276, pp. 459–466, Springer, Berlin,
Generation
Germany, 2005.
Figure 10: Convergence of the Team Plan cost for the 20-400 MAPP [6] A. Jacoff, E. Messina, and J. Evans, “Experiences in deploying
shown in Figure 9. test arenas for autonomous mobile robots,” in NIST Special
Publication, pp. 87–94, 2002.
References [7] S. Carpin, J. Wang, M. Lewis, A. Birk, and A. Jacoff, “High
fidelity tools for rescue robotics: results and perspectives,” in
[1] NASA, Mars Exploration Rover missions, January 2010, http:// RoboCup 2005: Robot Soccer World Cup IX, vol. 4020, pp. 301–
marsrovers.nasa.gov. 311, 2006.
Complexity 11

[8] G. Giardini and T. Kalmár-Nagy, “Centralized and distributed [27] E. Balas and P. Toth, “Branch and bound methods for the
path planning for multi-agent exploration,” in Proceedings of the traveling sales man problem,” Tech. Rep., DTIC Document,
AIAA Guidance, Navigation, and Control Conference, pp. 2701– 1985.
2712, Hilton Head Island, SC, USA, August 2007. [28] A. Schrijver, Theory of Linear and Integer Programming, John
[9] G. Giardini and T. Kalmár-Nagy, “Genetic algorithm for multi- Wiley & Sons, New York, NY, USA, 1986.
agent space exploration,” in Proceedings of the AIAA infoTech at [29] A. E. Carter, Design and application of genetic algorithms for
Aerospace Conference, vol. 2, pp. 1146–1160, May 2007. the multiple traveling salesperson assignment problem [Ph.D.
[10] S. Mitrovic-Minic and R. Krishnamutri, “The Multiple Trav- thesis], Department of Management Science and Information
eling Salesman Problem with Time Windows: bounds for the Technology, Virginia Polytechnic Institute and State University,
minimum number of vehicles,” Tech. Rep. TR 2002-11, FU CS 2003.
School, 2002.
[30] M. Kulich, J. Kubalik, J. Klema, and J. Faigl, “Rescue operation
[11] T. Bektas, “The multiple traveling salesman problem: an
planning by soft computing techniques,” in Proceedings of the
overview of formulations and solution procedures,” Omega, vol.
IEEE 4th International Conference on Intelligent Systems Design
34, no. 3, pp. 209–219, 2006.
and Application, pp. 103–109, Budapest, Hungary, 2004.
[12] S. Hong and M. W. Padberg, “A note on the symmetric multiple
traveling salesman problem with fixed charges,” Operations [31] S. Gorenstein, “Printing press scheduling for multi-edition
Research, vol. 25, no. 5, pp. 871–874, 1977. periodicals,” Management Science, vol. 16, no. 6, pp. B-373–B-
383, 1970.
[13] A. Singh and A. S. Baghel, “A new grouping genetic algorithm
approach to the multiple traveling salesperson problem,” Soft [32] L. Tang, J. Liu, A. Rong, and Z. Yang, “A multiple traveling
Computing, vol. 13, no. 1, pp. 95–101, 2009. salesman problem model for hot rolling scheduling in Shanghai
[14] D. S. Johnson and L. A. McGeoch, “The Traveling Salesman Baoshan Iron and Steel Complex,” European Journal of Opera-
Problem: a case study in local optimization,” in Local Search in tional Research, vol. 124, no. 2, pp. 267–282, 2000.
Combinatorial Optimization, pp. 215–310, 1997. [33] B. Gavish and K. Srikanth, “An optimal solution method for
[15] GaTech, Traveling Salesman Problem, January 2010, http:// large-scale multiple traveling salesmen problems,” Operations
www.tsp.gatech.edu. Research, vol. 34, no. 5, pp. 698–717, 1986.
[16] G. Gutin and A. Punnen, The Traveling Salesman Problem and [34] “First international conference on innovative computing, infor-
Its Variations, vol. 12 of Combinatorial Optimizations, Kluwer mation and control—TOC,” in Proceedings of the First Inter-
Academic, Norwell, Mass, USA, 2002. national Conference on Innovative Computing, Information and
[17] A. E. Carter and C. T. Ragsdale, “Scheduling pre-printed news- Control—Volume I (ICICIC ’06), vol. 3, August 2006.
paper advertising inserts using genetic algorithms,” Omega, vol. [35] I. Kara and T. Bektas, “Integer linear programming formu-
30, no. 6, pp. 415–421, 2002. lations of multiple salesman problems and its variations,”
[18] A. T. Stentz and B. Brummit, “GRAMMPS: a generalized European Journal of Operational Research, vol. 174, no. 3, pp.
mission planner for multiple mobile robots,” in Proceedings of 1449–1458, 2006.
the IEEE International Conference on Robotics and Automation, [36] C. S. Orloff, “Routing a fleet of 𝑀 vehicles to/from a central
May 1998. facility,” Networks, vol. 4, no. 2, pp. 147–162, 1974.
[19] A. T. Stentz and B. Brummit, “Dynamic mission planning for
[37] M. Bellmore and S. Hong, “Transformation of multisalesman
multiple mobile robots,” in Proceedings of the IEEE International
problem to the standard traveling salesman problem,” Journal
Conference on Robotics and Automation, April 1996.
of the Association for Computing Machinery, vol. 21, no. 3, pp.
[20] Z. Yu, L. Jinhai, G. Guochang, Z. Rubo, and Y. Haiyan, “An 500–504, 1974.
implementation of evolutionary computation for path planning
of cooperative mobile robots,” in Proceedings of the 4th World [38] G. Laporte, Y. Nobert, and S. Taillefer, “A branch-and-bound
Congress on Intelligent Control and Automation, vol. 3, pp. 1798– algorithm for the asymmetrical distance-constrained vehicle
1802, Shanghai, China, June 2002. routing problem,” Mathematical Modelling, vol. 9, no. 12, pp.
857–868, 1987.
[21] UMA, The VRP Web, January 2009, http://neo.lcc.uma.es/radi-
aeb/WebVRP. [39] A. I. Ali and J. L. Kennington, “The asymmetric M-travelling
[22] F. B. Pereira, J. Tavares, P. Machado, and E. Costa, “GVR: a salesman problem: a duality based branch-and-bound algo-
new genetic representation for the vehicle routing problem,” in rithm,” Discrete Applied Mathematics, vol. 13, no. 2-3, pp. 259–
Proceedings of the 13th Irish Conference on Artificial Intelligence 276, 1986.
and Cognitive Science (AICS ’02), pp. 95–102, Limerick Ireland, [40] A. S. Rostami, F. Mohanna, H. Keshavarz, and A. A. Hos-
September 2002. seinabadi, “Solving multiple traveling salesman problem using
[23] J. A. Bondy and U. S. R. Murty, Graph Theory with Applications, the gravitational emulation local search algorithm,” Applied
Macmillan, London, UK, 1976. Mathematics & Information Sciences, vol. 9, no. 2, pp. 699–709,
[24] R. Diestel, Graph Theory, Springer, 2005. 2015.
[25] G. Giardini and T. Kalmár-Nagy, “Genetic algorithm for com- [41] D. Sofge, A. Schultz, and K. De Jong, “Evolutionary compu-
binatorial path planning: the subtour problem,” Mathematical tational approaches to Solving the Multiple Traveling Sales-
Problems in Engineering, vol. 2011, Article ID 483643, 31 pages, man Problem using a neighborhood attractor schema,” in
2011. Applications of Evolutionary Computing, vol. 2279, pp. 153–162,
[26] S. Tschoke, R. Luling, and B. Monien, “Solving the traveling Springer, 2002.
salesman problem with a distributed branch-and-bound algo- [42] B. Na, Heuristics for no-depot multiple traveling salesmen prob-
rithm on a 1024 processor network,” in Proceedings of the IEEE lem with minmax objective [M.S. thesis], H. Milton School of
9th International Parallel Processing Symposium, pp. 182–189, Industrial and Engineering, Georgia Institute of Technology,
Santa Barbara, Calif, USA, April 1995. Atlanta, Ga, USA, May 2006.
12 Complexity

[43] A. E. Carter and C. T. Ragsdale, “A new approach to solving

the multiple traveling salesperson problem using genetic algo-
rithms,” European Journal of Operational Research, vol. 175, no.
1, pp. 246–257, 2006.
[44] E. C. Brown, C. T. Ragsdale, and A. E. Carter, “A group-
ing genetic algorithm for the multiple traveling salesperson
problem,” International Journal of Information Technology and
Decision Making, vol. 6, no. 2, pp. 333–347, 2007.
[45] M. T. Shirafkan, H. Seidgar, J. Rezaian-Zeidi, and N. Javadian,
“Using a hybrid simulated annealing and genetic algorithms for
non fixed destination multi-depot multiple traveling sales men
problem with time window and waiting penalty,” The Journal of
Mathematics and Computer Science, vol. 4, pp. 428–435, 2012.
[46] R. I. Bolaños, M. G. Echeverry, and J. W. Escobar, “A multi-
objective non-dominated sorting genetic algorithm (NSGA-II)
for the multiple traveling salesman problem,” Decision Science
Letters, vol. 4, no. 4, pp. 559–568, 2015.
[47] R. Kaliaperumal, A. Ramalingam, and J. Sripriya, “A modi-
fied Two part chromosome Crossover for solving mtsp using
Genetic algorithms,” in Proceedings of the International Confer-
ence on Advanced Research in Computer Science Engineering and
Technology (ICARCSET ’15), Unnao, India, March 2015.
[48] K. Sundar and S. Rathinam, “An exact algorithm for a hetero-
geneous, multiple depot, multiple traveling salesman problem,”
in Proceedings of the International Conference on Unmanned
Aircraft Systems (ICUAS ’15), pp. 366–371, Denver, Colo, USA,
June 2015.
[49] G. Sharon, R. Stern, M. Goldenberg, and A. Felner, “The
increasing cost tree search for optimal multi-agent pathfinding,”
Artificial Intelligence, vol. 195, pp. 470–495, 2013.
[50] D. E. Goldberg, Genetic Algorithms in Search, Optimization, and
Machine Learning, Addison-Wesley Professional, Boston, Mass,
USA, 1989.
[51] K. Bryant, Genetic algorithms and the traveling salesman problem
[Ph.D. thesis], Department of Mathematics, Harvey Mudd
College, Claremont, Calif, USA, 2000.
[52] K. Katayama, H. Sakamoto, and H. Narihisa, “The efficiency of
hybrid mutation genetic algorithm for the travelling salesman
problem,” Mathematical and Computer Modelling, vol. 31, no.
10–12, pp. 197–203, 2000.
[53] M. Matayoshi, M. Nakamura, and H. Miyagi, “A genetic
algorithm with the improved 2-opt method,” in Proceedings
of the IEEE International Conference on Systems, Man and
Cybernetics (SMC ’04), vol. 4, pp. 3652–3658, IEEE, The Hague,
The Netherlands, October 2004.
[54] J. L. Bentley, “Experiments on traveling salesman heuristics,” in
Proceedings of the 1st Annual ACM-SIAM Symposium on Dis-
crete Algorithms, pp. 91–99, Society for Industrial and Applied
Mathematics, San Francisco, Calif, USA, January 1990.
[55] H. Sengoku and I. Yoshihara, “A fast TSP solver using GA on
JAVA,” in Proceedings of the 3rd International Symposium on
Artificial Life, and Robotics (AROB III ’98), pp. 283–288, 1998.
[56] G. Reinelt, “TSPLIB: a traveling salesman problem library,”
ORSA Journal on Computing, vol. 3, no. 4, pp. 376–384, 1991.
[57] Multiple Traveling Salesmen Problem—Genetic Algorithm,
March 2008, http://www.mathworks.com/matlabcentral/fileex-
change/.
Advances in Advances in Journal of Journal of
Operations Research
Hindawi Publishing Corporation
Decision Sciences
Hindawi Publishing Corporation
Applied Mathematics
Hindawi Publishing Corporation
Algebra
Hindawi Publishing Corporation
Probability and Statistics
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

The Scientific International Journal of

World Journal
Hindawi Publishing Corporation
Differential Equations
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Submit your manuscripts at

https://www.hindawi.com

International Journal of Advances in

Combinatorics
Hindawi Publishing Corporation
Mathematical Physics
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal of Journal of Mathematical Problems Abstract and Discrete Dynamics in

Complex Analysis
Hindawi Publishing Corporation
Mathematics
Hindawi Publishing Corporation
in Engineering
Hindawi Publishing Corporation
Applied Analysis
Hindawi Publishing Corporation
Nature and Society
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

International
Journal of Journal of
Mathematics and
Mathematical
Discrete Mathematics
Sciences

Journal of International Journal of Journal of

Hindawi Publishing Corporation Hindawi Publishing Corporation Volume 2014

Function Spaces
Hindawi Publishing Corporation
Stochastic Analysis
Hindawi Publishing Corporation
Optimization
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014