Combinatorial Optimization for All: Using LLMs to Aid Non-Experts (journal.iberamia.org)

Recent research explores the use of Large Language Models (LLMs) to enhance optimization algorithms, specifically targeting non-experts. Unlike previous efforts focused on generating optimization algorithms from scratch, this study assesses whether LLMs can refine existing codebases for a variety of algorithms, including metaheuristics and reinforcement learning, when applied to the classic Traveling Salesman Problem. Impressive results emerged, as the methodology led to improved algorithm variants in 9 out of 10 cases, with LLMs autonomously implementing advanced techniques like heuristic initializations that significantly reduced runtime. This research holds notable significance for the AI/ML community by democratizing access to optimization techniques, making them more approachable for individuals without specialized training. The LLM-generated code maintained a high quality of software standards, with a maintainability index averaging 53.40 and a substantial reduction in cyclomatic complexity by 19.4% in certain models. These findings demonstrate not only the potential for LLMs to boost algorithmic performance but also their capability to streamline the development process, ultimately empowering a broader range of users to leverage advanced optimization strategies effectively.