Performance Evaluation of Population Seeding Techniques of Permutation-Coded GA Traveling Salesman Problems Based Assessment

Genetic algorithms (GAs) are a population-based meta-heuristic global optimization technique for dealing with complex problems with a very large search space. The population initialization is a crucial task in GAs because it plays a vital role in the convergence speed, problem search space exploration, and also the quality of the final optimal solution. Though the importance of deciding problem-specific population initialization in GA is widely recognized, it is hardly addressed in the literature. In this article, different population seeding techniques for permutation-coded genetic algorithms such as random, nearest neighbor (NN), gene bank (GB), sorted population (SP), and selective initialization (SI), along with three newly proposed ordered-distance-vector-based initialization techniques have been extensively studied. The ability of each population seeding technique has been examined in terms of a set of performance criteria, such as computation time, convergence rate, error rate, average convergence, convergence diversity, nearest-neighbor ratio, average distinct solutions and distribution of individuals. One of the famous combinatorial hard problems of the traveling salesman problem (TSP) is being chosen as the testbed and the experiments are performed on large-sized benchmark TSP instances obtained from standard TSPLIB. The scope of the experiments in this article is limited to the initialization phase of the GA and this restricted scope helps to assess the performance of the population seeding techniques in their intended phase alone. The experimentation analyses are carried out using statistical tools to claim the unique performance characteristic of each population seeding techniques and best performing techniques are identified based on the assessment criteria defined and the nature of the application.