An adaptive multitask optimization algorithm based on competitive scoring

Zhen Yang, Yiping Zhu, Yunliang Jiang, Yaochu Jin, Feng Ju, Yang Feng

Abstract: Evolutionary multitask optimization (EMTO) generally relies on transferring knowledge among different tasks to improve performance. However, how to reduce the impact of negative transfer in EMTO still remains a challenge, especially for complex many-task optimization problems. To address this issue, this article proposes an adaptive multitask optimization algorithm by introducing a competitive scoring mechanism (MTCS). Based on this mechanism, MTCS quantifies the effects of transfer evolution and self-evolution, and then adaptively sets the probability of knowledge transfer and selects source tasks. Then, a dislocation transfer strategy is designed in knowledge transfer to maximize the effects of evolution. MTCS is compared with ten state-of-the-art EMTO algorithms on multitask and many-task benchmark problems. The experimental results demonstrate the effectiveness of the proposed knowledge transfer strategy and the superiority of overall performance of MTCS.