Backtracking Ant System For The Traveling Salesman Problem

Backtracking Ant System
for the Traveling Salesman Problem
Sameh Al-Shihabi
Industrial Engineering Department, University of Jordan, Amman 11942, Jordan

shihabi sameh@hotmail.com
Abstract. In this work, we adopt the concept of backtracking from

the Nested Partition (NP) algorithm and apply it to the Max-Min Ant
System (MMAS) to solve the Traveling Salesman Problem (TSP). A new
type of ants that is called backtracking ants (BA) is used to challenge a
subset of the solution feasible space that is expected to have the global
optimum solution. The size of this subset is decreased if the BAs find
a better solution out of this subset or increased if the BAs fail in their
challenge. The BAs don’t have to generate full tours like previous ant
systems, which leads to a considerable reduction in the computation
effort. A computational experiment is conducted to check the validity of
the proposed algorithm.
1 Introduction
The Traveling Salesman Problem (TSP) consists of a number of nodes that are
fully connected by edges and it is required to visit all of these nodes, return to
the starting city and avoid the formation of sub tours. An arc between city i and
city j is penalized an amount dij and the objective of the TSP is to minimize
the sum of these penalties[4].
Ant Colony Optimization (ACO) algorithms imitate the natural behavior of
ants in finding the shortest distance between their nests and food sources. Ants
exchange information about good routes through a chemical substance called
pheromone that accumulates for short routes and evaporate for long routes. The
first ACO algorithm is the Ant system (AS) [2] that did not show good results
with respect to solving large TSPs.
Different modification to the original AS are introduced to improve the qual-
ity of the solution such as having local and global updating of the pheromone
matrix as in the Ant Colony system (ACS) [3] or allowing certain ants only
to update the pheromone matrix [2] . The Max-Min Ant System (MMAS)[8]
modifies the AS by keeping the pheromone values within a range [τmax , τmin ]
to ensure that there is an upper bound to the probability that a certain edge is
selected [8].
This work adopts a number of concepts from the Nested Partition (NP)
algorithm and adds it to the MMAS. The Nested partition (NP) [6, 5] algorithm
is based on adaptive sampling where the algorithm moves from a subregion to
another in the feasible space rather than moving between solution points. At
M. Dorigo et al. (Eds.): ANTS 2004, LNCS 3172, pp. 318–325, 2004.

c Springer-Verlag Berlin Heidelberg 2004
Backtracking Ant System for the Traveling Salesman Problem 319
each step of the NP method, the most promising subregion is divided into M
subregions while the abandoned subregions are aggregated to form a surrounding
subregion. These M + 1 subregions are sampled and a promising index is used to
choose which subregion is the most promising to be partitioned further, hence,
more closely inspected. The algorithm backtracks to a larger subregion if the
surrounding subregion is found to be the most promising. The algorithm stops
once it reaches a subregion that cannot be partitioned further.
This work presents a new type of AS that called Backtracking Ant System
(BAS). Similar to the NP algorithm, the BAS moves from a subset of the feasible
region to another. The proposed algorithm uses a new type of ant called Back-
tracking Ant (BA) to find new routes, in a manner similar to the sampling of the
surrounding subregion in the NP algorithm. If a BA finds a shorter tour than
the incumbent one, the algorithm backtracks to a larger subregion containing
the new sample. The BAS is tested in this work by solving the TSP.
A hybrid between the MMAS and NP is already implemented in [1]. The
sampling scheme implemented in [1] uses ants for finding some samples then
finds other samples by perturbing the ants’ tours. This work is different from [1]
since it does not implement the partitioning step and does not sample the sub-
regions resulting from partitioning the most promising subregion. Additionally,
all samples are found by ants in this work.
2 BAS Algorithm
Due to the large number of symbols used to describe the proposed algorithm,
this section begins by summarizing the notation used. We try to keep the same
MMAS notation used in [8] and some of the NP notation used in [6]. The BAS
algorithm consists of a number of steps and these steps are explained using an
assumed problem instance having 5 cities that are numbered as 0, 1, ..., 4.
k The iteration number.

Nc The number of cities of the TSP instance.
dij The distance between city i and city j.
Lg (k) The value of the best solution found by iteration k.
Tg (k) A vector showing the best tour found by iteration k.
Tc (k) A vector of variable size that shows the subset of edges challenged by the
backtracking ants.
d(k) The depth of iteration k which represent the number of edges challenged
by the backtracking ants.
m Number of backtracking ants used to challenge Tc (k).
τij The pheromone value between city i and j.
τmax (k) The maximum value of the pheromone at iteration k.
τmin (k) The minimum value of the pheromone at iteration k.
ρ The evaporation rate of the pheromone.
ηij A heuristic information which is taken as d1.0
ij
for the TSP.
α A constant indicating the importance of the pheromone value.
β A constant Indicating the importance of the heuristic information.

Backtracking Ant System For The Traveling Salesman Problem

Uploaded by

Copyright:

Available Formats

Backtracking Ant System For The Traveling Salesman Problem

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Backtracking Ant System For The Traveling Salesman Problem

Uploaded by

Copyright:

Available Formats

Backtracking Ant System

for the Traveling Salesman Problem

Industrial Engineering Department, University of Jordan, Amman 11942, Jordan

Abstract. In this work, we adopt the concept of backtracking from

k The iteration number.

You might also like