| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 4.71 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Os problemas de dimensionamento de lotes têm como objetivo identificar as quantidades ideais de produção para atender à procura, minimizando, simultaneamente, os custos de setup e o custo de stock. Esta dissertação explora como os LLMs está transformou a forma como o problema de dimensionamento de lotes é abordado. O trabalho foca-se inicialmente na evolução da literatura sobre técnicas de IA aplicadas a problemas de dimensionamento de lotes, com o intuito de fornecer uma visão abrangente da área de estudo, focando depois nas plataformas de LLMs destacando as diversas formas de aplicação, e mais tarde, no problema de dimensionamento de lotes. Apesar do interesse crescente da comunidade científica na aplicação do ChatGPT, a sua utilização em problemas de otimização ainda não tem recebido atenção. A variabilidade intrínseca ao ChatGPT torna a engenharia de prompt essencial para adaptar a ferramenta às especificidades de cada problema, contudo, é frequentemente negligenciada. Esta realidade destaca a necessidade de desenvolver uma abordagem estruturada para observar, testar e avaliar a aplicabilidade do ChatGPT em problemas de dimensionamento de lotes. Desta forma foi investigada a eficácia das plataformas ChatGPT e
Copilot na resolução do problema de dimensionamento de lotes. Inicialmente, ambas as plataformas apresentaram resultados insatisfatórios, com inconsistências e falta de repetibilidade. Para superar estas limitações, foi testado o modelo GPT-4 do ChatGPT, mais tarde substituído pelo GPT-4o, que demonstrou maior consistência e capacidade de captar o contexto do problema, resultando em soluções próximas do ótimo. Após várias iterações, foi desenvolvido um prompt otimizado que proporcionou resultados mais consistentes na resolução do problema com o ChatGPT. Este prompt foi posteriormente aplicado em duas
amostras distintas para avaliar o desempenho da LLMs m diferentes cenários. Esta abordagem permitiu examinar a eficácia e consistência do ChatGPT em condições variadas, oferecendo uma análise mais robusta da sua aplicabilidade e adaptabilidade aos desafios específicos de cada amostra. Os resultados mostraram que, em ambas as amostras, o ChatGPT superou as restantes heurísticas analisadas, não apresentando diferenças estatisticamente significativas em relação ao método de Wagner-Whitin.
Lot-sizing problems aim to identify the optimal production quantities to meet demand while simultaneously minimizing setup and inventory costs. This dissertation explores how Large Language Models (LLMs) have transformed the approach to lot-sizing problems. The work initially focuses on the evolution of literature of AI techniques applied to lot-sizing problems, aiming to provide a comprehensive view of the study area, then shifts to LLM platforms, highlighting various application methods, and later delves into the lot-sizing problem itself. Despite the growing interest of the scientific community in applying ChatGPT, its use in optimization problems has yet to receive significant attention. The intrinsic variability of ChatGPT makes prompt engineering essential for adapting the tool to the specifics of each problem, however, this is often overlooked. This reality underscores the need to develop a structured approach to observe, test, and evaluate the applicability of ChatGPT in lot-sizing problems. Thus, the effectiveness of ChatGPT and Copilot platforms in solving batch sizing problems was investigated. Initially, both platforms presented unsatisfactory results, with inconsistencies and a lack of repeatability. To overcome these limitations, the GPT-4 model of ChatGPT was tested and later replaced by GPT-4o, which demonstrated greater consistency and ability to capture the context of the problem, resulting in near-optimal solutions. After several iterations, an optimized prompt was developed, which provided more consistent results in solving the problem with ChatGPT. This prompt was subsequently applied to two distinct samples to assess the performance of LLMs in different scenarios. This approach allowed examining the effectiveness and consistency of ChatGPT under varied conditions, offering a more robust analysis of its applicability and adaptability to the specific challenges of each sample. The results showed that in both samples, ChatGPT outperformed the other heuristics analyzed, with no statistically significant differences compared to the Wagner-Whitin method.
Lot-sizing problems aim to identify the optimal production quantities to meet demand while simultaneously minimizing setup and inventory costs. This dissertation explores how Large Language Models (LLMs) have transformed the approach to lot-sizing problems. The work initially focuses on the evolution of literature of AI techniques applied to lot-sizing problems, aiming to provide a comprehensive view of the study area, then shifts to LLM platforms, highlighting various application methods, and later delves into the lot-sizing problem itself. Despite the growing interest of the scientific community in applying ChatGPT, its use in optimization problems has yet to receive significant attention. The intrinsic variability of ChatGPT makes prompt engineering essential for adapting the tool to the specifics of each problem, however, this is often overlooked. This reality underscores the need to develop a structured approach to observe, test, and evaluate the applicability of ChatGPT in lot-sizing problems. Thus, the effectiveness of ChatGPT and Copilot platforms in solving batch sizing problems was investigated. Initially, both platforms presented unsatisfactory results, with inconsistencies and a lack of repeatability. To overcome these limitations, the GPT-4 model of ChatGPT was tested and later replaced by GPT-4o, which demonstrated greater consistency and ability to capture the context of the problem, resulting in near-optimal solutions. After several iterations, an optimized prompt was developed, which provided more consistent results in solving the problem with ChatGPT. This prompt was subsequently applied to two distinct samples to assess the performance of LLMs in different scenarios. This approach allowed examining the effectiveness and consistency of ChatGPT under varied conditions, offering a more robust analysis of its applicability and adaptability to the specific challenges of each sample. The results showed that in both samples, ChatGPT outperformed the other heuristics analyzed, with no statistically significant differences compared to the Wagner-Whitin method.
Description
Keywords
Lot-Sizing Artificial intelligence LLMs ChatGPT Heuristics Prompt Dimensionamento de lotes Inteligência artificial Heurísticas