[2E1-GS-5b] Agents

10:00 AM - 10:15 AM JST(1:00 AM - 1:15 AM UTC)

[2E1-GS-5b-05] An LLM-Based Automatic Optimization Framework with Abstract and Solver-Specific Formulations

〇Kuniaki Satori¹ (1. Mitsubishi Electric Corporation)

Keywords:

Autonomous Optimization,Problem Formulation,Algorithm Selection,LLM,Optimization

Optimization is an important technology used in many engineering areas such as design, operation, and planning. However, building mathematical models and choosing suitable algorithms require expert knowledge, which makes optimization difficult for non-experts. Recently, large language models (LLMs) have been studied to automatically solve optimization problems written in natural language. However, many previous studies fix the mathematical model at an early stage, which strongly limits the choice of solution methods. This can cause mismatches between the model and the solver. In this study, we propose a new framework for automatic optimization generation based on three steps: abstract formulation, method selection, and method-specific formulation. The framework first represents the problem structure in an abstract way, then selects an appropriate solving strategy, and finally rebuilds a mathematical model that fits the chosen method. Experimental results show improvements in both solution quality and execution success rate compared with previous approaches, suggesting that the proposed framework is effective for real-world optimization tasks.