Dynamic Multiobjective Optimization Aided by ESN-based Prediction Approach

Dynamic multi-objective problems (DMOPs) have aroused extensive attention in recent years. Prediction-based methods have been proven to be effective. However, most existing methods assume the linear relationships between historical solutions. For real-life systems, ignoring the complex nonlinear relationships between historical environments may result in low prediction accuracy. To solve this problem, the echo state network (ESN) based prediction approach is proposed for DMOPs. First, the reservoir of ESN is used to express the input dynamics of the historical solutions to explore the linear or nonlinear relationships among historical solutions. Then, a fractal interpolation technique (FIT) is introduced to enrich the training data while preserving the original time series features as much as possible. The final experimental results show that the designed algorithm can solve the dynamic multi-objective optimization problems effectively.