LLM-Enhanced Accumulated Learning Prediction for Evolutionary Dynamic Multiobjective Optimization

Prediction-based dynamic multiobjective evolutionary algorithms (DMOEAs) often fail to fully exploit comprehensive historical information when solving dynamic multiobjective optimization problems (DMOPs). To alleviate this challenge, this paper proposes a DMOEA with a large language model (LLM)-enhanced accumulated learning prediction for addressing DMOPs. Specifically, a bidirectional long short-term memory (BiLSTM) network is trained online using approximated Pareto-optimal solutions from all previously encountered environments, enabling it to effectively capture the problem’s dynamics and maintain a favorable trade-off between convergence and diversity. Moreover, as a novel contribution to this field, an LLM is used to refine the BiLSTM model by generating high-quality solutions as enhanced training data, thereby substantially reducing prediction error and enhancing robustness. Comprehensive empirical studies have been conducted on 14 commonly used benchmark DMOPs and a real-world problem, which demonstrate the superior performance of the proposed algorithm over five competitive DMOEAs. © 2026 IEEE.