Multiobjective Multitasking Optimization With Decomposition-Based Adaptive Knowledge Transfer
提出一种多目标多任务进化算法，通过分解框架和自适应子问题选择来调整任务间的知识迁移比例，并设计自适应迁移策略选择最合适的迁移算子，在标准测试集和神经网络架构搜索中验证了有效性。
Multiobjective multitasking optimization (MTO) is an emerging research direction in the evolutionary computation community, which tries to solve multiple optimization problems concurrently by utilizing shared search knowledge among related tasks. However, most existing algorithms of MTO achieve the knowledge transfer without quantifying the differences among tasks and ignore the differences in the characteristics of transfer operators, which may degrade the convergence speed. To alleviate this issue, this article proposes a multiobjective multitasking evolutionary algorithm with decomposition-based adaptive knowledge transfer (MMTEA-DAKT). Specifically, an adaptive subproblems selection method is designed, which adopts a decomposition-based framework to decompose the MTO problem into a series of single-objective optimization subproblems, aiming to adjust the proportion of knowledge transfer among all different tasks based on the improvement rate of each subproblem. Besides, an adaptive knowledge transfer strategy is devised to select the most appropriate knowledge transfer operator, which aims to improve the efficiency of knowledge transfer. To verify the effectiveness of our proposed MMTEA-DAKT, we compare it with several advanced related algorithms on three standard multiobjective multitasking test suites and the practical application of neural architecture search. The experimental results show that MMTEA-DAKT has a significant competitive advantage in solving most of the problems compared to several state-of-the-art algorithms.