Multitasking optimization via an adaptive solver multitasking evolutionary framework

Abstract

Multitasking optimization (MTO) aims to solve multiple tasks simultaneously in a single run. Many multitasking evolutionary algorithms (MTEAs) have been developed in recent years for solving MTO problems. Existing MTEAs typically use a single solver to handle multiple optimization tasks. However, different tasks have distinct characteristics, such as convex, nonconvex, and multimodal. If one could automatically find a best-fitting solver for each task, it would be more efficient to solve different tasks. To this end, we propose a multitasking evolutionary framework based on adaptive solver selection, namely MTEA-SaO, where a suitable knowledge transfer strategy is embedded. The proposed method can be featured as 1) It explicitly assigns several solver subpopulations to each task and adaptively picks a best-fitting solver for each task, and 2) It enables knowledge transfer using the implicit similarities between tasks. The experimental results have demonstrated the effectiveness of solver adaptation and knowledge transfer strategies and the overall superior performance of MTEA-SaO.