Evolutionary Multitasking Optimization With Dual-Space Similarity Assisted Knowledge Transfer

Evolutionary multitasking optimization (EMTO) aims to enhance the overall efficiency of optimizing multiple tasks through inter-task knowledge transfer, with its performance heavily dependent on task similarity. Prevailing domain adaptation methods try to align the distribution of some high-quality solutions in the decision space, yet consider little about their fitness landscape consistency, potentially resulting in negative transfer. To address this issue, this study proposes a new EMTO algorithm with a dual-space similarity assisted knowledge transfer. After aligning the task-specific subspaces through a PCA-based linear transformation, the dual-space similarity metric further computes the Spearman rank correlation coefficient between the fitness rankings of some representative solutions in source task and their counterparts in target task, thereby explicitly quantifying fitness landscape consistency. Based on the dual-space similarity, an adaptive knowledge transfer strategy is designed. It regulates the transfer probability according to the absolute value of the similarity metric, while selects the solution to be transferred considering the metric sign. Consequently, even a non-elite source solution may facilitate positive transfer. Comprehensive experiments conducted on multiple benchmarks and a real-world problem demonstrate the efficacy of the proposed algorithm.