Enhancing Knowledge Transfer in the EMTO with Manifold Learning and Reinforcement Learning

Evolutionary multitask optimization (EMTO) algorithms are a new paradigm in the field of evolutionary algorithm. They can simultaneously handle two or more optimization problems in a single run. Knowledge transfer is the core of this algorithmic framework. Although existing EMTO algorithms are working hard to alleviate the impact of negative transfer, their solution performance remains unsatisfactory, especially when the similarity between tasks is low or the number of tasks increases. To enhance the transfer effect between tasks and reduce the occurrence of negative transfer by differences in feature ratios and angles, we propose an EMTO algorithm based on manifold learning and reinforcement learning within a multi-population evolutionary framework, named MPEF-MRL. The main contributions of our paper are as follows: 1) Use maximized density divergence (MDD) to measure the similarity between the different tasks. 2) Apply manifold learning and CORAL algorithm to reduce the difference in features and distributions between the different tasks. 3) Use the Q-learning algorithm to adjust the relevant parameters. In our experiments, we validate the proposed MPEF-MRL algorithm using the CEC2017, WCCI2020-MTSO and WCCI2020-MaTSO benchmark suite. The experimental results demonstrate the effectiveness of the knowledge transfer strategy and the superior performance of the algorithm in overall performance.