In manufacturing sector, efficiently scheduling jobs is crucial due to global competition and the need for rapid product delivery, cost reduction, and operational efficiency. This thesis explores how large language models (LLMs) can automate the job-shop scheduling problem (JSP) formulations, aiming to make the scheduling process more agile and efficient in dynamic manufacturing environments.

The study tested four different methods for generating JSP problem formulations using LLMs. The first method involved zero-shot examples with GPT-4, where the model relied entirely on pre-existing knowledge. The second method used one-shot examples with GPT4, providing the model with one specific pre-defined or random example to guide its responses. The third method employed few-shot examples with GPT-4 using Retrieval-Augmented Generation (RAG), where the model dynamically selected a few relevant examplesto inform its output. The fourth method involved fine-tuning a smaller model, Google gemma-2B, on a problem-formulation dataset to improve its performance.

The methods were evaluated through automated and manual analyses. An automated Python script validated the problem formulations and checked for syntax correctness. Additionally, a manual review was conducted to further analyse the results, errors and assess overall performance.

The results showed significant differences in effectiveness. The zero-shot GPT-4 method had only 5% valid formulations, with many syntax errors. The one-shot GPT-4 method improved to 55% valid formulations, reducing syntax issues. The few-shot GPT-4 with RAG method achieved the highest success rate with 90% valid formulations, demonstrating the benefits of using contextually relevant examples. In contrast to expectation, the fine-tuned gemma-2B model failed to generate any valid formulations, highlighting its limitations in handling complex tasks even after fine-tuning.

The study concludes that incorporating domain-specific knowledge and examples with language models can improve automated problem formulation, thus optimizing the job scheduling processes, making them both faster and more adaptable to changing conditions. However, the study also observed limitation with smaller models like gemma-2B and recommends further research into fine-tuning techniques and alternative specialized code generating models. Overall, integrating domain-specific knowledge and contextual examples enhances the quality and accuracy of automated problem formulations, offering promising improvements for scheduling in manufacturing and other fields.