AGLSDC: A Genetic Local Search Suitable for Parallel Computation

Abstract

Because evolutionary algorithms (EAs) generally require many repeated evaluations of objective functions, it often takes considerable time to solve optimization problems. Parallel computation is one means to shorten the required computation time. In earlier works, the authors proposed an EA suitable for coarse-grained parallel computers, a genetic local search with distance independent diversity control (GLSDC). Though GLSDC has been applied successfully to several practical problems, its parallel efficiency abruptly drops off as the number of CPUs for computation increases. To achieve a higher parallel efficiency, the authors now propose a new EA, an asynchronous GLSDC (AGLSDC), constructed by reworking the algorithm of GLSDC. This paper introduces the proposed method and reports verification of the method through numerical experiments on several benchmark problems and a practical problem.