グループ化と爆発によるPSOの効率化

抄録

In practical engineering problems, there are more and more requirements for efficient method that can solve large scale problems. In this study, large scale means the number of design variables are more than 15. Particle Swarm Optimization (PSO) is one of choices that we do not need sensitivity and also it works well even in multi-peaks problems. However, its efficiency is only limited to small scale problems. Most likely number of design variables are less than 6. When the number of design variables raises there are more possibilities to fall into local optimum and PSO cannot get out from there. By its nature, if one achieved to become global best, its velocity becomes close to zero and fall into local optimum. In this study, Agent PSO is used as basic. Agent PSO means that each parameter settings have different characteristics, such as early convergence, middle convergence and late convergence. To avoid early convergence situations, usually we need a new global best to have velocity. In this study, grouping is proposed. It means that we divide population into several groups. Each group has its own global best. And when there are no improvements of personal bests for several iterations, it will look up the other groups global. As for explosions, they are something like mutation in Genetic Algorithms. In this study, two different ways are proposed. If velocities of individuals those who are early convergence, we gave specific number of design variables to random numbers. In several generations, they tend to get close to global bests for their groups and have a chance to get out from local optimum. Another explosion is for those groups which had worst global best, some individuals in that group are given randomly between global best and itself. Effectiveness of the proposed method are given through both Rosenbrock function and Rastrigin function for large scale problems such as 128, 64, 32 and 16 design variables to less than 1.0e-11 order.