Large Language Model Driven Evolutionary Optimization Benchmark Generation Algorithm
Yuhiro Ono, Tomohiro Harada, Yukiya Miura · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2025
Many optimization benchmarks have been proposed for single and multi-objective problems to assess the performance of optimization algorithms, including evolutionary algorithms. They are manually generated or developed based on real-world scenarios. However, manually generated benchmarks often fail to capture essential properties of real-world problems. On the other hand, real-world-based benchmarks require an extensive run time or high cost of creating benchmarks and are sometimes inaccessible due to confidential restrictions. To overcome these limitations, this study proposes a large language model-driven evolutionary optimization benchmark generator (LLM-EBG) that automatically generates benchmarks with desired characteristics. LLM-EBG integrates an evolutionary algorithm with a large language model (LLM). Candidate benchmarks are generated through crossover and mutation using an LLM and refined through an evolutionary process. As an initial attempt, we generated unconstrained single-objective benchmarks designed to emphasize the difference between genetic algorithms (GA) and differential evolution (DE). Experimental results demonstrated that the proposed method can generate benchmarks where the performance gap between GA and DE is significant; in particular, GA outperforms DE, and vice versa. These results highlight the potential of the proposed method to create tailored and informative benchmarks for optimization algorithm evaluation.