Loading...
Publication
Aplicação de modelos de linguagem de grande escala a problemas de escalonamento das operações
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 4.92 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Os problemas de Escalonamento das Operações são particularmente relevantes em ambientes industriais, e fundamentais para tornar a produção eficiente e cumprir as datas estipuladas com os clientes. Em ambientes de grande dimensão, encontrar soluções ótimas pode tornar-se um desafio que requer tempo e recursos computacionais significativos. No entanto, o rápido avanço da Inteligência Artificial tem sido uma enorme ajuda no ramo da gestão industrial. Um dos principais campos de IA a ser desenvolvido nos últimos anos é o de IA Generativa. Neste trabalho procura-se explorar as capacidades de três Modelos de Linguagem de Grande Escala
distintos, DeepSeek R1, GPT o4-mini e GPT o3, através da resolução de um conjunto de 19 instâncias em ambiente Job Shop de diferente complexidade. Os modelos demonstraram variabilidade no desempenho, reforçando a importância do aperfeiçoamento dos prompts introduzidos através de engenharia de prompts. Para ajudar a estabilizar os resultados, foi desenvolvido um manual de prompting com foco na resolução deste tipo de problemas, independentemente do modelo a ser utilizado. Os dois primeiros modelos não conseguiram
gerar soluções para mais de metade das instâncias, sendo observada uma dificuldade em respeitar as restrições impostas nos problemas mais complexos. O modelo DeepSeek R1, em especial, apenas conseguiu gerar solução para 6 das 19 instâncias e apresentou a pior eficiência entre os três modelos, demorando por vezes mais de 10 minutos a resolver problemas de menor complexidade. Já o GPT o3, disponível numa plataforma de subscrição, demonstrou um desempenho superior, gerando soluções ótimas ou próximas do ótimo para todas as instâncias,
embora ainda tenha precisado de um tempo de processamento não negligenciável para gerar as respostas. Concluiu-se que, com o cuidado adequado e preparação prévia dos prompts, este modelo representa um recurso relevante na otimização dos problemas de Escalonamento das Operações. Adicionalmente, como o modelo GPT o4-mini apresentou um tempo de processamento das respostas muito inferior aos outros dois, pode constituir-se como uma alternativa eficiente para a otimização de problemas menos complexos.
Operations Scheduling problems are particularly relevant in industrial environments and are fundamental to ensuring efficient production and meeting client deadlines. In large-scale environments, finding optimal solutions can be a challenge that requires significant time and computational resources. However, the rapid advancement of Artificial Intelligence has been a great help in the field of industrial management. One of the main areas of AI being developed in recent years is Generative AI. This work seeks to explore the capabilities of three distinct Large Language Models, DeepSeek R1, GPT o4-mini and GPT o3, by solving a set of 19 instances of varying complexity in Job Shop environment. The models showed some variability in their performance, reinforcing the importance of improving the prompts introduced, through the implementation of prompt engineering. To help stabilize the results, a manual of prompting was developed focusing on solving this type of problem, regardless of the model being used. The first two models were unable to generate solutions for more than half of the instances, with difficulty observed in respecting the constraints imposed on the most complex problems. The DeepSeek R1 model was only able to generate solutions for 6 of the 19 instances and had the worst efficiency among the three models, sometimes taking more than 10 minutes to solve less complex problems. GPT o3, available on a subscription platform, demonstrated superior performance, generating optimal or near-optimal solutions for all instances, although it still required a significant amount of processing time to generate the responses. It was concluded that, with proper care and prior preparation of the prompts, this model represents a relevant resource in the optimization of Operations Scheduling problems. Additionally, as the GPT o4-mini model had a much shorter response processing time than the other two, it may be an efficient alternative for optimizing less complex problems.
Operations Scheduling problems are particularly relevant in industrial environments and are fundamental to ensuring efficient production and meeting client deadlines. In large-scale environments, finding optimal solutions can be a challenge that requires significant time and computational resources. However, the rapid advancement of Artificial Intelligence has been a great help in the field of industrial management. One of the main areas of AI being developed in recent years is Generative AI. This work seeks to explore the capabilities of three distinct Large Language Models, DeepSeek R1, GPT o4-mini and GPT o3, by solving a set of 19 instances of varying complexity in Job Shop environment. The models showed some variability in their performance, reinforcing the importance of improving the prompts introduced, through the implementation of prompt engineering. To help stabilize the results, a manual of prompting was developed focusing on solving this type of problem, regardless of the model being used. The first two models were unable to generate solutions for more than half of the instances, with difficulty observed in respecting the constraints imposed on the most complex problems. The DeepSeek R1 model was only able to generate solutions for 6 of the 19 instances and had the worst efficiency among the three models, sometimes taking more than 10 minutes to solve less complex problems. GPT o3, available on a subscription platform, demonstrated superior performance, generating optimal or near-optimal solutions for all instances, although it still required a significant amount of processing time to generate the responses. It was concluded that, with proper care and prior preparation of the prompts, this model represents a relevant resource in the optimization of Operations Scheduling problems. Additionally, as the GPT o4-mini model had a much shorter response processing time than the other two, it may be an efficient alternative for optimizing less complex problems.
Description
Keywords
Operations Scheduling Artificial Intelligence Large Language Models ChatGPT DeepSeek Prompting Escalonamento das operações Inteligência artificial Modelos de Linguagem de Grande Escala