An Extended ISM for Globally Multimodal Function Optimization by Genetic Algorithms

When attempting to optimize a function where exists several big-valley structures, conventional GAs often fail to find the global optimum. Innately Split Model (ISM) is a framework of GAs, which is designed to avoid this phenomenon called UV-Phenomenon. However, ISM doesn't care about previously-searched areas by the past populations. Thus, it is possible that populations of ISM waste evaluation cost for redundant searches reaching previously-found optima. In this paper, we introduce Extended ISM (EISM) that uses search information of past populations as trap to suppress overlapping searches. To show performance of EISM, we apply it to some test functions, and analyze the behavior.