Sequential and Parallel Genetic Algorithms for solving Travelling Salesman Problem

Blaž Poje (2013) Sequential and Parallel Genetic Algorithms for solving Travelling Salesman Problem. EngD thesis.

Preview

PDF Download (1665Kb)

Abstract

The main aim of this thesis is the comparison of parallel and sequential algorithm implementation for solving Travelling Salesman Problem using Genetic algorithms. The parallel algorithm is being executed on a graphical processing unit and sequential algorithm on central processing unit. The latter algorithm has been converted into a parallel form by using processor threads. It also focuses on the quality of the developed solutions in co-dependence with different sets of crossovers and mutations with the presence of migrations. Contrary to our expectations, the parallel algorithm does not reach speedups on the graphical processing unit. The analysis demonstrates the commonly used crossovers to be unfit for implementation on massively parallel architectures. We have not been successful in transforming them to parallel form. We have successfully proven the positive effects of migrations on general solution fitness. The smaller the number of solutions being transferred between separate populations, the bigger the effect. Distribution of average solution fitness could offer further guidelines for selection of crossovers that could be beneficial if transformed into parallel form.

Actions (login required)