In this thesis we consider the statistical properties of the symmetric travelling
salesman problem (TSP). Previous work on the statistical properties of the problem
has been largely limited to the Euclidean case with vertex coordinates as random
variables with known distribution embedded in Rd, and to the case of independent
identically distributed random edge costs. Furthermore, this previous work did not
extend to computing the moments, beyond the mean. In the work presented here
we consider the more general case of problem instances specified as a set of edge
costs and with no (known) embedding or coordinate system available.
For an instance of the problem on n vertices with fixed edge costs we give
constructive proofs that the population variance of tour costs over the solution
space can be computed in O(n2), the third central moment can be computed in
O(n4) and the fourth central moment can be computed in O(n6). These results
provide direct methods to compute the moments about the origin and factorial
moments of these orders with the corresponding computational complexity. In
addition the results provide tractable methods to compute, among other statistics,
the standard deviation of tour costs, the skewness of the probability distributions
of tour costs over the solution space and kurtosis of this distribution.
In the case of the stochastic TSP with edge costs defined as independently
distributed random variables with (not necessarily identical) known mean and
variance we provide a O(n4) algorithm to compute the variance of tour costs.
Given a subgraph S of a tour in an n city TSP, we provide an O(n2) algorithm
to compute the expected tour costs over the solution space of those tours containing
S. This is useful in analysing and constructing algorithms such as Gutin's greedy
expectation heuristic.
We demonstrate that the probability distribution of gains over the 2-opt landscape of an n city TSP can be computed in O(n4 log(n)). This result provides
a tractable algorithm to compute, among other statistics the moments of gains
over the landscape. The result also provides the 2-opt neutrality (the number of
neighbouring solutions with identical cost) of a instance. The result has natural
generalisation to the 3-opt landscape (at higher computational complexity). We
relate the variance of tour costs over the solution space to that of the gains over
the 2-opt landscape of a problem, providing an O(n2) method to compute the
variance of gains over the landscape.
We apply our method to compute the low order moments of the distribution of
tour costs to several empirical studies of the solution space. Among other results
we: con¯rm the known relationship between the standard deviation of tour costs
and the optimal tour cost; we show a correlation between the skewness and the
optimal tour cost; we demonstrate that the moments can be used to estimate the
complete probability distribution of tour costs.
Description:
University of Technology, Sydney. Faculty of Engineering and Information Technology.