Knee Point Based Imbalanced Transfer Learning for Dynamic Multi-objective Optimization

Contribution

Uses knee-point-aware transfer to avoid negative transfer when fronts shift unevenly.

Abstract

When transferring knowledge across optimization tasks, the source and target problems rarely overlap uniformly — the source may concentrate high-quality solutions near one region of the Pareto front while the target demands coverage elsewhere. Blindly transferring all source solutions can therefore introduce negative transfer, slowing the search. This paper proposes a knee-point-based imbalanced transfer framework that detects when distributions are mismatched and selectively transfers solutions from regions of the Pareto front where transfer is most likely to help. The result is robust performance gains across a range of dynamic multiobjective benchmarks with varying degrees of distribution shift.

Why This Paper Matters

A clean diagnosis of when transfer learning helps versus hurts, plus a principled fix. It is the standard reference when arguing for selective rather than unconditional transfer in evolutionary optimization.

BibTeX

@article{jiang2020knee,
  title={Knee Point Based Imbalanced Transfer Learning for Dynamic Multi-objective Optimization},
  author={JIANG, Min and WANG, Zhenzhong and HONG, Haokai and G. YEN, Gary},
  journal={IEEE Transactions on Evolutionary Computation},
  year={2020},
}