Multi-View Transfer with Marginal and Conditional Alignment for Many-Task Optimization
Sheng-Hao Wu, Jian-Yu Li, Zhi‐Hui Zhan, Jun Zhang · 2023
Many-task optimization problem (MaTOP) is a kind of problem in that more than three optimization tasks are required to be solved simultaneously. How to utilize multiple different kinds of similarities among many tasks for knowledge transfer (KT) is a challenging issue in solving MaTOP. To this end, this paper proposes a multi-view transfer (MVT)-based differential evolution (MVTDE) for effectively transferring knowledge in solving MaTOP. In particular, a perspective of task alignment is regarded as a view of performing KT. By combining different views including marginal and conditional distribution alignment, MVT can better fully utilize various task similarities from different aspects to solve MaTOP efficiently. This paper is with two contributions. First, we propose the MVT method that incorporates multiple KT methods from the views of marginal and conditional distribution alignment to enhance the search efficiency on the target tasks. Second, we propose to control the occurrence of different KT methods in MVT according to probability parameters and propose a parameter update strategy to adaptively adjust the probability parameters according to the feedback of different KT methods. In the experimental study, we carry out comparative experiments on three widely used MaTOP benchmarks to validate the effectiveness of MVTDE. The results show that MVTDE obtains competitive performance compared to several state-of-the-art many-task optimization algorithms.