Skip to main navigation Skip to search Skip to main content

Bi-stage learning adaptive multi-task optimization

  • Chu-Xiang LI
  • , Zi-Jia WANG*
  • , Qian YUE*
  • , Qian-Hui ZHANG
  • , Chang-Long WANG
  • , Zhi-Hui ZHAN
  • , Yu-Song TAN
  • , Hua LI
  • , Sam KWONG
  • , Jun ZHANG
  • *Corresponding author for this work

Research output: Journal Publications › Journal Article (refereed) › peer-review

Abstract

Evolutionary multi-task optimization (EMTO) is an emerging research topic in the evolutionary computation (EC) community, which tackles multiple tasks simultaneously by inter-task knowledge transfer (KT) and intra-task self-evolution (SE). Therefore, how to balance the inter-task KT and the intra-task SE remains the main challenge in the field of EMTO. In this paper, we propose a bi-stage learning adaptive multi-task optimization (BLAMTO) with two learning stages: the pre-learning KT stage and the post-learning SE stage, for solving multi-task optimization problems (MTOPs). Specifically, in the pre-learning KT stage, a simple and effective mixed population transfer (MPT) strategy is designed to mix individuals from different populations and preserve the promising offspring to promote KT process. In the post-learning SE stage, an elitist local search (ELS) strategy is developed to further refine the solution accuracy. To achieve balance between the inter-task KT and the intra-task SE, an adaptive stage switch (ASS) strategy is designed to adaptively switch the stage from the pre-learning KT stage to the post-learning SE stage by evaluating inter-task KT performance. The experimental results on MTOP benchmarks CEC17 and CEC22 indicate that BLAMTO generally achieves competitive or superior performance compared to other state-of-the-art EMTO algorithms, demonstrating the overall effectiveness of BLAMTO. Furthermore, BLAMTO is applied to a real-world multi-task planar kinematic arm control (PKAC) application, manifesting its applicability. Finally, experiments on the many-task optimization problem (MaTOP) benchmarks CEC19 and WCCI20 further illustrate its scalability.

Original languageEnglish
Article number116655
JournalApplied Soft Computing
Volume204
Early online date20 Oct 2026
DOIs
Publication statusE-pub ahead of print - 20 Oct 2026

Bibliographical note

Publisher Copyright:
© 2026 Elsevier B.V.

Funding

This work was supported in part by the National Natural Science Foundations of China (NSFC) under Grants 62106055 and 62394333, in part by the Guangdong Natural Science Foundation under Grant 2025A1515010256, 2026A1515011523, and 2026A1515060004, and in part by the Guangzhou Science and Technology Planning Project under Grants 2023A04J0388 and 2023A03J0662.

Keywords

  • Evolutionary multitask optimization (EMTO)
  • Bi-stage strategy
  • Mixed transfer strategy
  • Adaptive stage switch

Fingerprint

Dive into the research topics of 'Bi-stage learning adaptive multi-task optimization'. Together they form a unique fingerprint.

Cite this