A mixture-of-experts prediction framework for evolutionary dynamic multiobjective optimization

Contribution

Uses a mixture of predictors to match different change patterns in dynamic optimization.

Abstract

Dynamic optimization problems vary in how predictable their changes are: some change smoothly and almost linearly; others change abruptly with rare, large jumps. A single prediction model cannot serve both regimes well. This paper proposes a mixture-of-experts prediction framework for evolutionary dynamic multiobjective optimization that learns multiple specialized predictors, each handling a particular change regime, and adaptively weights them based on the current environment. The framework consistently improves convergence on benchmarks that mix gradual and abrupt changes.

Why This Paper Matters

A modular and extensible framework for handling heterogeneous change dynamics — useful as a template for any predictive-transfer approach to non-stationary optimization.

BibTeX

@article{rambabu2019mixture,
  title={A mixture-of-experts prediction framework for evolutionary dynamic multiobjective optimization},
  author={Rambabu, Rethnaraj and Vadakkepat, Prahlad and Tan, Kay Chen and Jiang, Min},
  journal={IEEE transactions on cybernetics},
  volume={50},
  number={12},
  pages={5099--5112},
  year={2019},
  publisher={IEEE},
}