LLM-Enhanced Adaptive MOEA/D for Solving Multiobjective Problems with Complex Pareto Fronts

Abstract

Adaptively adjusting the components or parameters of decomposition-based multiobjective evolutionary algorithms (MOEA/Ds) is crucial for enhancing its performance in handling multiobjective optimization problems with complex Pareto fronts. However, component design strategies and parameter tuning mechanisms based on domain knowledge remain the primary bottleneck limiting performance improvement. This paper proposes a large language model (LLM)-enhanced MOEA/D algorithm. The algorithm leverages LLM’s domain priors and prompt engineering techniques to automatically generate modular code, replacing manually designed search operators. Furthermore, by reasoning about the algorithm’s iterative process information, it achieves adaptive dynamic adjustments of the aggregation function, weight strategy, and parameter rules. Experimental results on 12 benchmark test instances with complex Pareto fronts demonstrate that the proposed LLM-driven MOEA/D variant exhibits significant advantages over traditional methods in the HV performance metric.