Individual-based transfer learning for dynamic multiobjective optimization

Contribution

Transfers promising individuals instead of whole populations, enabling finer-grained adaptation.

Abstract

Most transfer-learning approaches for dynamic optimization transfer entire populations at once, treating individuals identically regardless of their role in the search. This paper argues that individual solutions carry individual learning signals — some are stable landmarks, others are transient explorers — and proposes an individual-based transfer framework that scores and selects which individuals to transfer based on their predicted future usefulness. The approach outperforms population-level transfer on diverse DMOP benchmarks and provides a finer-grained mechanism for balancing exploration and exploitation across environments.

Why This Paper Matters

A clean conceptual move — from populations to individuals — that has been picked up by later work in evolutionary multi-task learning and surrogate-assisted optimization.

BibTeX

@article{jiang2020individual,
  title={Individual-based transfer learning for dynamic multiobjective optimization},
  author={Jiang, Min and Wang, Zhenzhong and Guo, Shihui and Gao, Xing and Tan, Kay Chen},
  journal={IEEE Transactions on Cybernetics},
  volume={51},
  number={10},
  pages={4968--4981},
  year={2020},
  publisher={IEEE},
}