Randomness Selection in Differential Evolution Using Thompson Sampling
Akira Notsu, Junya Tsubamoto, Yuichi Miyahira, Seiki Ubukata, Katsuhiro Honda · 2020
Differential evolution is a versatile and highperformance search algorithm for optimization problems. Some algorithms adaptively change the strategy of differential evolution, and they have better performance depending on the setting of the strategy. We have developed a parameter-independent algorithm by simply rethinking the setting of these strategies in terms of randomness. Furthermore, if the adaptation process of strategies and randomness can be regarded as a Bandit problem, more efficient adaptation process can be developed. In this study, we consider adapting several randomness patterns to differential evolution using Thompson sampling, which is a bandit algorithm with very good performance. The effectiveness was confirmed by numerical experiments.