Automatic Algorithm Design for Operations Optimization Using Large Language Models and Evolutionary Techniques

The business background highlights the limitations of traditional operations‑optimization algorithms, which rely heavily on expert knowledge, have long development cycles, and often fail to achieve global optimality, leading to low resource utilization and high operational costs.

Large models offer powerful generation and optimization capabilities, enabling automatic creation of customized algorithms by learning from massive data and optimization cases, and by integrating evolutionary concepts to avoid local optima and accelerate deployment.

Section 2 describes the automatic algorithm design workflow, identifying pain points such as low development efficiency, poor adaptability, and limited optimization effects. It introduces genetic algorithms (GA) as a robust global‑search method and explains how large models can generate high‑quality initial solutions, guide population initialization, and dynamically adjust strategies during evolution.

The reflection mechanism is detailed, comprising short‑term and long‑term feedback loops that compare code performance, generate improvement suggestions, and refine the code through dynamic execution using exec , fitness evaluation, and iterative population updates.

Implementation steps cover code generation via LLM APIs with prompt engineering, code extraction, dynamic execution, timeout handling, module import resolution, population initialization, fitness scoring, and optimization through crossover, mutation, and elite selection.

Section 2.5 applies the framework to personnel‑allocation recommendation in logistics, defining input data, model objectives, and output structures, and outlining heuristic modeling, data preprocessing, constraint formulation, and solution verification.

The future outlook (Section 3) proposes broader applications such as automated test‑code generation, algorithm optimization, AI‑assisted programming assistants, code knowledge bases, vertical domain models, simulation systems, data visualization, and online learning for continuous optimization, emphasizing the scalability and impact of LLM‑driven automated algorithm design.