MIXTURE-OF-EXPERTS BASED PREDICTION AND ADAPTIVE POPULATION REINITIALIZATION - ACCELERATING CONVERGENCE IN EVOLUTIONARY DYNAMIC MULTIOBJECTIVE OPTIMIZATION

Abstract

Multiobjective optimization involves simultaneous optimization of two or more objective functions that are conflicting in nature, which results in a set of trade-off solutions for a given optimization problem. Real-world optimization problems can have objective functions that may change with time. Dynamic multiobjective optimization requires evolutionary algorithms to detect changes and robustly track the varying Pareto-optimal solutions. The thesis proposes a mixture-of-experts (MoE) framework, a novel ensemble scoring scheme (ESS), and an adaptive linear Kalman filter (ALKF) for population prediction to accelerate convergence in DMOEAs. The ensemble frameworks, MoE and ESS, combine multiple population reinitialization mechanisms to improve the quality of the population reinitialized after every detected change. The ALKF, MoE, and ESS are integrated with popular MOEAs and are evaluated on 13 popular benchmark functions against state-of-the-art DMOEAs. The experimental results show that the proposed algorithms significantly improve the dynamic optimization performance compared to the state-of-the-art DMOEAs.