Adaptive operator selection via deep reinforcement learning for decomposition-based multi-objective multi-task optimization

Abstract Multi-objective multi-task optimization (MO-MTO), which aims to simultaneously solve multiple related multi-objective problems by leveraging knowledge transfer across tasks to enhance optimization performance on each task, has emerged as a new research focus in the field of evolutionary computation. Although a variety of MO-MTO algorithms have achieved certain success on different problems, most of them rely on a single genetic operator to generate offspring, and their knowledge transfer process is often confined to the objective space, failing to fully exploit valuable information in the decision space. This limitation results in insufficient search efficiency and difficulty in producing high-quality offspring. To address this issue, this paper proposes a multi-task evolutionary algorithm with reinforcement learning-based knowledge selection (MTEA-RLKS), which integrates a collaborative knowledge transfer mechanism combining both objective and decision spaces, as well as static and dynamic knowledge. Specifically, the algorithm extracts static knowledge from global population distributions and local neighborhoods in the objective space, while capturing dynamic knowledge of population evolution in the decision space by employing Gaussian processes. Through synergistic knowledge transfer, it guides the generation of high-quality offspring and accelerates the optimization process. Additionally, a Deep Q-Network is introduced to adaptively select genetic operators, which is trained online during iterations to accommodate the evolutionary needs of different solutions. Experimental results on two MO-MTO benchmark test sets, CEC2017 and CEC2019, demonstrate that the proposed MTEA-RLKS significantly outperforms six other state-of-the-art algorithms.