Using LLM for automatic evolvement of metaheuristics from swarm algorithm SOMA

Abstract

This study investigates the use of the GPT-4 Turbo, a large language model, to enhance the Self-Organizing Migrating Algorithm (SOMA), specifically its All to All variant (SOMA-ATA). Utilizing the model's extensive context capacity for iterative prompting without feedback, we sought to autonomously generate superior algorithmic versions. Contrary to our initial hypothesis, the improvements did not progress linearly. Nevertheless, one iteration stood out significantly, consistently outperforming the baseline across various pairwise comparisons and showing a robust performance profile. This iteration's structure deviated substantially from traditional SOMA principles, underscoring the potential of large language models to create distinctive and effective algorithmic strategies. The results affirm the methodology's ability to produce high-performing algorithms without expert intervention, setting the stage for future research to integrate feedback mechanisms and conduct detailed code analyses to understand further the modifications made by the Large Language Models.