Requirement-Driven Adaptive Differential Evolution for Many-Task Optimization

Abstract

The major challenge in evolutionary many-task optimization (EMaTO) is to transfer knowledge among the tasks effectively. Moving on existing EMaTO algorithms, which mainly focus on what kind of knowledge to select and how to transfer the selected knowledge, this article focuses on a novel perspective to tackle the issue of when to transfer knowledge. This gives rise to a requirement-driven adaptive differential evolution (RADE) algorithm. Specifically, the RADE introduces a dual-scale evolutionary state estimator (ESE) to evaluate the state of evolution from both temporal and spatial scales, obtaining the real optimization requirement of the population and thus adaptively controlling the timing of knowledge transfer. This way, RADE not only leads to efficiently dealing with the when issue, but also drives two adaptive strategies for knowledge selection and knowledge transfer, leading to more efficient dealing with the what and the how issues, respectively. In particular, a requirement-driven adaptive knowledge selection strategy merges the evolutionary state similarity and the population distribution similarity for enhancing the selection of relevant source knowledge. Then, a requirement-driven adaptive knowledge transfer strategy is executed based on adaptive parameters and operators, thus significantly improving the efficiency of knowledge transfer. Experimental results on the CEC19 and WCCI22 benchmarks show that RADE outperforms the compared state-of-the-art EMaTO algorithms in both effectiveness and efficiency.