Transfer Learning-Based Dynamic Multiobjective Optimization Algorithms

Contribution

Introduces transfer learning for dynamic multiobjective optimization, so search can reuse prior environments.

Abstract

Dynamic multiobjective optimization problems (DMOPs) arise whenever objectives, constraints, or environmental conditions change over time — from dynamic vehicle routing to online portfolio optimization. Conventional evolutionary algorithms restart their search after each change, wasting knowledge accumulated in earlier environments. This paper introduces a transfer-learning framework that explicitly reuses historical populations, decision-variable distributions, and Pareto-set shapes to bootstrap the search after a change is detected. The result is markedly faster convergence to the new Pareto front than from-scratch restarts, establishing transfer learning as a foundational technique for dynamic evolutionary optimization and inspiring much of the subsequent work in this area.

Why This Paper Matters

This paper opened an entire research direction — transfer learning for evolutionary dynamic optimization — and is the most-cited work in the field. Reading it gives you the foundational framework that most later algorithms build on.

BibTeX

@article{jiang2017transfer,
  title={Transfer Learning-Based Dynamic Multiobjective Optimization Algorithms},
  author={Jiang, Min and Huang, Zhongqiang and Qiu, Liming and Huang, Wenzhen and Yen, Gary G.},
  journal={IEEE Transactions on Evolutionary Computation},
  volume={22},
  number={4},
  pages={501--514},
  year={2018},
  doi={10.1109/tevc.2017.2771451},
  publisher={IEEE},
}