LAIGA: LLM-assisted Interactive Genetic Algorithm for Preference-incorporated Multi-objective Optimization

Abstract

Large Language models (LLMs) have demonstrated strong logical reasoning capabilities in evolutionary computation, particularly in generating new solutions, automatically designing algorithms, and evaluating the quality of solutions. Interactive evolutionary computation (IEC) is an approach that allows the decision maker (DM) to intervene in the algorithm execution process to influence the direction of evolution. In traditional interactive evolutionary processes, it is usually necessary to model the DM’s preferences so that the algorithm can understand these preferences and guide the iterative process on behalf of the DM, thereby reducing the DM’s workload. Although many studies have adopted different methods to construct preference models based on various types of preference information, whether LLM can serve as an auxiliary model in IEC remains unexplored. Therefore, this paper proposes a new algorithm, LAIGA, which is based on the genetic algorithm (GA) and integrates an LLM-assisted module to allow the DM to provide preferences in natural language form. The LLM then assists the algorithm in automatically adjusting the weight vector of the preference model. Finally, we designed experiments to evaluate the algorithm’s performance on the DTLZ2 problem in dimensions 6–12 under three different types of DM preference models, confirming its effectiveness. The experimental results show that the proposed method achieves performance comparable to mainstream approaches, while reducing the number of DM interactions and cognitive burden, and ensuring that the final solution obtained is sufficiently satisfactory to the DM. This work provides new insights into the application of LLM in the field of IEC.