所属 広島修道大学 商学部 職種 教授
|A Study on Multi-Armed Bandit Algorithms for Dynamic Selection of Parameters and Topologies in Particle Swarm Optimization
|Operations Researche and Information Systems
|Kyushu University Press
|Setsuko Sakai and Tetsuyuki Takahama
|The multi-armed bandit problem is deﬁned as maximizing the total reward when the reward of each choice is unknown and a choice is sequentially selected from multiple choices. Particle Swarm Optimization (PSO) has been successfully applied to various optimization problems. Various parameter (including topology) settings are known for PSO, but it is difﬁcult to determine the appropriate setting because the setting depends on the problem to be solved and the search process. In this study, we propose to apply bandit algorithms to the parameter setting. If a new position after a movement is better than the personal best position found so far, the reward is 1 as success, otherwise the reward is 0 as failure. The setting that maximizes the cumulative reward is discovered by the bandit algorithms. The effectiveness of the proposed method is shown by introducing the method to PSO and optimizing 13 benchmark problems.