A Mixture-of-Experts Prediction Framework for Evolutionary Dynamic Multiobjective Optimization.

Area of Science:

• Computational Intelligence
• Optimization Theory
• Machine Learning

Background:

• Dynamic multiobjective optimization (DMOO) necessitates tracking evolving Pareto-optimal solutions (POS) in dynamic environments.
• Existing prediction mechanisms use past population data to accelerate convergence to true POS upon environmental change.
• Robust POS prediction remains a challenge in DMOO.

Purpose of the Study:

• To propose a novel mixture-of-experts (MoE)-based ensemble framework for robust prediction of Pareto-optimal solutions (POS) in dynamic multiobjective optimization.
• To enhance the accuracy and efficiency of POS tracking in changing environments.
• To improve the overall performance of dynamic optimization algorithms.

Main Methods:

• An ensemble framework utilizing multiple prediction mechanisms is developed.
• A gating network dynamically switches between predictors based on their performance over time.
• The framework is evaluated on 13 benchmark dynamic multiobjective optimization problems.

Main Results:

• The proposed MoE framework significantly improves dynamic optimization performance.
• Performance gains are particularly notable for problems with distinct dynamic POS shifts in decision space.
• The framework demonstrates effectiveness in handling problems with highly nonlinear decision variable linkages.

Conclusions:

• The mixture-of-experts ensemble framework offers a robust approach to POS prediction in dynamic multiobjective optimization.
• This method enhances the ability to track evolving optimal solutions in complex, changing environments.
• The findings suggest a significant advancement in dynamic optimization strategies, especially for challenging problem types.