Abstract:

The manual transition from natural language descriptions to formal mathematical models is a critical yet underappreciated bottleneck in operations research. This paper evaluates the feasibility of using Large Language Models (LLMs), specifically ChatGPT and Gemini, to automate the formulation of the complex Scheduling-Location (ScheLoc) problem. Unlike existing literature that often focuses on basic optimization tasks, this study utilizes five diverse, real-world scenarios ranging from mobile diagnostic units to humanitarian food relief to test model performance. The evaluation employed two query strategies: one assessing independent problem recognition and another providing the specific “ScheLoc” designation as a prompt. Our findings reveal that while LLMs struggle to independently identify the specific problem nomenclature, they demonstrate a robust functional understanding by achieving 100% accuracy in identifying optimization criteria across all tested cases. Furthermore, providing the problem name significantly improved the models’ ability to extract correct variables, domains, and parameters. The results demonstrate that LLMs can effectively handle the heavy lifting of data extraction and initial mathematical formulation, allowing human experts to transition from primary modelers to high-level verifiers. This approach offers a feasible path toward streamlining complex modeling workflows and democratizing expert-level optimization for industries lacking specialized operations research staff.