解空間の特徴と自己組織化マップを利用した探索手法

Abstract

Function optimizations are used in various fields. Solution spaces, however, varies according to the solving problems. Therefore we think that searching algorithms considering solution spaces are more efficient than algorithms without considering. Our proposed algorithm uses self-organizing maps (SOM) to realize the pseudo solution space. From this pseudo solution space in SOM map, we presume candidate optimum solutions. We obtain the “number feature” and the “soluiton feature” from these candidate optimum solutions. The “number feature” means how many candidate optimum solutions are in the SOM map. The “soluiton feature” means what kind of features the candidate optimum solutions have. We prepare six features as these “soluiton feature”s. Prepared features are “Dimension”, “Surround”, “Difference”, “Vector”, “Fourier” and “Wavelet”. We compare our proposed algorithm with particle swarm optimization, differential evolution and SPX. The problem we treat in experiments is the real-valued minimization problem with several benchmark functions. We show that our proposed algorithm searches the optimum solution effectively in terms of the number of function evaluations from experimental results.