An Online Machine Learning-Based Prediction Strategy for Dynamic Evolutionary Multi-objective Optimization

摘要

Due to the impact of environmental changes, dynamic evolutionary multi-objective optimization algorithms need to track the time-varying Pareto optimal solution set of dynamic multi-objective optimization problems (DMOPs) as soon as possible by effectively mining historical data. Since online machine learning can help algorithms dynamically adapt to new patterns in the data in machine learning community, this paper introduces Passive-Aggressive Regression (PAR, a common online learning technology) into dynamic evolutionary multi-objective optimization research area. Specifically, a PAR-based prediction strategy is proposed to predict the new Pareto optimal solution set of the next environment. Furthermore, we integrate the proposed prediction strategy into the multi-objective evolutionary algorithm based on decomposition with a differential evolution operator (MOEA/D-DE) to handle DMOPs. Finally, the proposed prediction strategy is compared with three state-of-the-art prediction strategies under the same dynamic MOEA/D-DE framework on CEC2018 dynamic optimization competition problems. The experimental results indicate that the PAR-based prediction strategy is promising for dealing with DMOPs.