Loading...
Publication
Sistema de seleção e autoparametrização de meta-heurísticas com Machine Learning
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 6.16 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O escalonamento de produção apresenta como objetivo principal a definição de um plano que
atinja as metas de produção, garantindo que todas as restrições sejam cumpridas e otimizadas.
Consiste num processo que desempenha um papel fundamental nas empresas, identificandose
como um fator que influencia a eficiência operacional e qualidades das organizações. A
resolução de problemas de escalonamento é uma tarefa complexa e necessita da
implementação de metodologias adequadas. A principal motivação deste trabalho surgiu da
necessidade de oferecer uma contribuição para a resolução de problemas de escalonamento.
As Meta-heurísticas, caraterizadas pela sua fácil implementação, podem ser utilizadas para a
resolução de problemas de escalonamento. No entanto, a sua aplicação apresenta desafios,
principalmente na escolha da técnica mais adequada para a resolução de um determinado
problema. Além disso, as Meta-heurísticas possuem parâmetros que necessitam de ser
definidos de forma adequada.
As Meta-heurísticas e a Machine Learning são frequentemente combinadas com o intuito de
diminuir as suas desvantagens, melhor as suas capacidades e, com isso, melhorar o
desempenho do sistema de produção. As abordagens de Machine Learning podem ser
integradas com o intuito de auxiliar na seleção e parametrização das Meta-heurísticas.
Neste contexto, surge a necessidade de desenvolver um sistema de seleção e
autoparametrização de Meta-heurísticas que utiliza Machine Learning para otimizar o
escalonamento de tarefas em ambientes de produção. Este sistema abrange unicamente
problemas de escalonamento do tipo Job-Shop e a parametrização das Meta-heurísticas é
realizada de forma offline. O módulo considera apenas técnicas de aprendizagem
supervisionada de Machine Learning.
O sistema proposto foi estruturado em dois modelos preditivos: o primeiro destina-se à seleção
da Meta-heurística mais adequada para um problema de escalonamento; o segundo é
responsável pela afinação dos parâmetros da Meta-heurística selecionada. Os dois modelos
preditivos foram avaliados com recurso a indicadores de desempenho adequados, permitindo
a realização de uma análise detalhada da eficácia da integração das técnicas de Machine
Learning.
The main objective of production scheduling is to define a plan that achieves production targets, ensuring that all constraints are met and optimised. It is a process that plays a fundamental role in companies, identifying itself as a factor that influences the operational efficiency and quality of organisations. Solving scheduling problems is a complex task and requires the implementation of appropriate methodologies. The main motivation for this work arose from the need to contribute to the resolution of scheduling problems. Meta-heuristics, characterised by their easy implementation, can be used to solve scheduling problems. However, their application presents challenges, mainly in choosing the most appropriate technique for solving a given problem. In addition, Meta-heuristics have parameters that need to be defined appropriately. Meta-heuristics and Machine Learning are often combined to reduce their disadvantages, improve their capabilities and, thereby, improve the performance of the production system. Machine learning approaches can be integrated to assist in the selection and parameterisation of meta-heuristics. In this context, there is a need to develop a Meta-heuristic selection and self-parameterisation system that uses Machine Learning to optimise task scheduling in production environments. This system only covers Job-Shop scheduling problems, and the parameterisation of Metaheuristics is performed offline. The module only considers supervised Machine Learning techniques. The proposed system was structured into two predictive models: the first is intended for selecting the most appropriate Meta-heuristic for a scheduling problem; the second is responsible for fine-tuning the parameters of the selected Meta-heuristic. The two predictive models were evaluated using appropriate performance indicators, allowing for a detailed analysis of the effectiveness of integrating Machine Learning techniques.
The main objective of production scheduling is to define a plan that achieves production targets, ensuring that all constraints are met and optimised. It is a process that plays a fundamental role in companies, identifying itself as a factor that influences the operational efficiency and quality of organisations. Solving scheduling problems is a complex task and requires the implementation of appropriate methodologies. The main motivation for this work arose from the need to contribute to the resolution of scheduling problems. Meta-heuristics, characterised by their easy implementation, can be used to solve scheduling problems. However, their application presents challenges, mainly in choosing the most appropriate technique for solving a given problem. In addition, Meta-heuristics have parameters that need to be defined appropriately. Meta-heuristics and Machine Learning are often combined to reduce their disadvantages, improve their capabilities and, thereby, improve the performance of the production system. Machine learning approaches can be integrated to assist in the selection and parameterisation of meta-heuristics. In this context, there is a need to develop a Meta-heuristic selection and self-parameterisation system that uses Machine Learning to optimise task scheduling in production environments. This system only covers Job-Shop scheduling problems, and the parameterisation of Metaheuristics is performed offline. The module only considers supervised Machine Learning techniques. The proposed system was structured into two predictive models: the first is intended for selecting the most appropriate Meta-heuristic for a scheduling problem; the second is responsible for fine-tuning the parameters of the selected Meta-heuristic. The two predictive models were evaluated using appropriate performance indicators, allowing for a detailed analysis of the effectiveness of integrating Machine Learning techniques.
Description
Keywords
Scheduling Meta-heuristics Machine Learning Tuning Parameters Escalonamento Meta-heurísticas Afinação Parâmetros