Manifold Transfer Learning for Multitask Optimization

  • Wu, Ze-Bin; 
  • Wang, Zi-Jia; 
  • Zhan, Zhi-Hui; 
  • Tan, Yu-Song; 
  • Kwong, Sam; 
  • ... Zhang, Jun
Citations

WEB OF SCIENCE

2
Citations

SCOPUS

0

초록

Knowledge transfer (KT) plays an essential role in evolutionary multitask optimization (EMTO) to utilize knowledge from source tasks to facilitate the optimization of target tasks. However, most existing KT methods can only achieve superficial KT, typically by transferring elite or random individuals without deeply exploring the relationships between tasks, leading to ineffective transfer. To address this limitation, this article proposes a manifold-transfer-learning-based KT (MTLKT) strategy to deeply explore the similarities between tasks and effectively learn the mapping relationships for information prediction, thus achieving high-quality KT. Specifically, principal component analysis (PCA) is first applied to extract the low-dimensional feature subspaces. Next, a series of intermediate subspaces are sampled in the manifold space using the sampling geodesic flow (SGF) method to characterize the continuous change between the source and target populations. Finally, the transferred individual in the target population is gradually generated to approach the underlying geometric distribution of the superior individuals from the source population in the manifold space, improving KT efficiency. Experimental results on the multitask benchmarks, CEC2017 and CEC2022, demonstrate that manifold-transfer-learning-based multitask optimization (MTLMTO) significantly outperforms other state-of-the-art EMTO algorithms, achieving the best results on 16 out of all 38 tasks. Furthermore, experimental results on the many-task benchmarks CEC2019 and WCCI2020 highlight the scalability of MTLMTO. Moreover, MTLKT can be seamlessly integrated into other EMTO algorithms to further enhance their performance. Finally, MTLMTO achieves the best results across all 12 tasks of the multitask robot navigation application, demonstrating its applicability. © 2013 IEEE.

키워드

Evolutionary computation (EC); evolutionary multitask optimization (EMTO); knowledge transfer (KT); manifold transfer learning (MTL); SEARCH
제목
Manifold Transfer Learning for Multitask Optimization
저자
Wu, Ze-Bin; Wang, Zi-Jia; Zhan, Zhi-Hui; Tan, Yu-Song; Kwong, Sam; Zhang, Jun
DOI
10.1109/TCYB.2026.3738565
발행일
2026-10
유형
Article in press
저널명
IEEE Transactions on Cybernetics