Evolution strategies for a parallel multi-objective genetic algorithm
Ricardo Szmit, Amnon Barak · 2000
This paper compares evolution strategies for a parallel multi-objective genetic algorithm adopting the concept of Pareto optimality. This algorithm was applied to the solution of a set of process scheduling problems that are part of a standard scheduling benchmark. Our main goal was to compare the efficiency and the efficacy of the evolution strategies, and how they relate to attributes of the problem. In order to quantify the quality of populations produced by the algorithm, we measured the coverage of the solution space and the proximity to the Pareto-optimal front. Our results show that an evolution strategy using heterogeneous subpopulations with restart is consistently superior to traditional strategies, without being more expensive. We also observe that the performance of the algorithm is directly related to the problem's communication to computation ratio (CCR). Our approach is based on the division of the scheduling problem to two parts that can be solved independently. This division allows a simpler encoding of individuals, so that the crossover and mutation operations can be implemented more efficiently. Thanks to the combination of genetic search with proven heuristics, this gain in efficiency does not imply a loss of efficacy.