Dynamic multi-objective evolutionary algorithm based on dual-layer collaborative prediction under multiple perspective

Prediction-based strategies become increasingly prominent in addressing dynamic multi-objective optimization problems (DMOPs). However, challenges remain in selecting predictive models and effectively utilizing historical solutions. In this paper, we propose a multiple perspective dual-layer collaborative prediction strategy to efficiently tackle both challenges. The multi-perspective approach is further divided into a search perspective and a spatial perspective and realized through the collaboration of three sub-strategies. From the search perspective, we employ a dual-layer prediction strategy that focuses on both global and local information. Specifically, the first layer utilizes Gaussian process regression (GPR) to predict centrality, which serves as a measure of the population's collective intelligence. This layer effectively captures global insights into population dynamics, identifying overarching movement trends over time. Building on these global insights, the second layer employs a knee-point interval partitioning strategy that combines vector partitioning with knee-point-based predictions. This layer provides localized insights that complement the broader movement trends identified by the first layer. From the spatial perspective, we implement dual- layer historical similarity detection across non-dominated solutions in both decision and objective spaces. Specifically, the historical Pareto-similarity selection strategy identifies populations in these spaces that demonstrate the greatest similarity to the current population's non-dominated solutions. The spatial perspective complements the search perspective, forming a coherent framework that systematically integrates global, local, and historical information. Experimental results indicate that the proposed algorithm performs better than previous state-of-the-art methods.