Evolutionary Multitask Optimization Based on Search Behavior Learning

Evolutionary multitask optimization (EMTO) is an emerging topic in evolutionary computation to solve multitask optimization problems (MTOPs) with the help of knowledge transfer (KT). However, the current EMTO algorithms often ignore the information of the search behavior, which only exploit individual information or historical performance when estimating task similarity or transferring knowledge. In this paper, we propose a search behavior learning-based multitask optimization (SBLMTO) algorithm to address the challenge. The search behavior similarity (SBS) strategy is first designed to estimate the similarity of tasks based on the search behavior, so as to adaptively adjust the KT frequency. Then, the knowledge archive is introduced to store the useful transferred knowledge, further enhance the KT efficiency. Next, a two-level transferring knowledge learning (TTKL) strategy is proposed to achieve the effective learning from both the search behavior and the knowledge archive. SBLMTO is tested on the widely used multitask test suites CEC2017 and CEC2022, and the experimental results show that SBLMTO performs significantly better than other state-of-the-art EMTO algorithms, showing its superiority.