Técnicas de aprendizado de máquina para predição do custo da logística de transporte: uma aplicação em empresa do segmento de autopeças
Carregando...
Data
2020-03-02
Autores
Rodríguez, Elen Yanina Aguirre
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Em diferentes aspectos da vida cotidiana, o ser humano é forçado a escolher entre várias opções, esse processo é conhecido como tomada de decisão. No nível do negócio, a tomada de decisões desempenha um papel muito importante, porque dessas decisões depende o sucesso ou o fracasso das organizações. No entanto, em muitos casos, tomar decisões erradas pode gerar grandes custos. Desta forma, alguns dos problemas de tomada de decisão que um gerente enfrenta comumente são, por exemplo, a decisão para determinar um preço, a decisão de comprar ou fabricar, em problemas de logística, problemas de armazenamento, etc. Por outro lado, a coleta de dados tornou-se uma vantagem competitiva, pois pode ser utilizada para análise e extração de resultados significativos por meio da aplicação de diversas técnicas, como estatística, simulação, matemática, econometria e técnicas atuais, como aprendizagem de máquina para a criação de modelos preditivos. Além disso, há evidências na literatura de que a criação de modelos com técnicas de aprendizagem de máquina têm um impacto positivo na indústria e em diferentes áreas de pesquisa. Nesse contexto, o presente trabalho propõe o desenvolvimento de um modelo preditivo para tomada de decisão, usando as técnicas supervisionadas de aprendizado de máquina, e combinando o modelo gerado com as restrições pertencentes ao processo de otimização. O objetivo da proposta é treinar um modelo matemático com dados históricos de um processo decisório e obter os preditores compostos por funções empíricas que serão posteriormente utilizadas e modeladas de acordo com as restrições do problema. Assim, este trabalho pode ser classificado como uma pesquisa aplicada, com objetivos empíricos descritivos e experiência prática que explicarão o modelo e suas vantagens. A maneira de abordar o problema deste trabalho será quantitativa, sendo os procedimentos técnicos de modelagem e simulação. A sistemática proposta é validada aplicando-se a um problema real em uma empresa multinacional brasileira do segmento de autopeças, situado no Vale do Paraíba-SP. O conjunto de dados vem do processo de transporte, visando aplicar a sistemática e criar um modelo matemático para prever o custo de transporte de uma quantidade de itens para um destino específico, de modo que este modelo finalmente ajuda a reduzir os tempos de estimativa e tomar melhores decisões. Finalmente, para a criação do modelo preditivo, foram testadas as técnicas de aprendizagem supervisionada de regressão linear, árvore de decisão CART, árvore de modelo M5P, máquina de vetor de suporte e redes neurais. Todos os modelos gerados foram avaliados por as métricas estatísticas de desempenho do erro médio absoluto, raiz do erro quadrático médio, erro absoluto relativo, erro quadrático relativo, coeficiente de correlação e coeficiente de determinação ou também conhecido como R2. Sendo que, o modelo M5P obteve os menores valores nas métricas de erro e os maiores valores na correlação e R2, mostrando ser um modelo eficiente, além de fornecer as equações necessárias para o processo de otimização.
In different aspects of everyday life, the human being is forced to choose between several options, this process is known as decision making. At the business level, decision making plays a very important role, because the success or failure of organizations depends on these decisions. However, in many cases, making the wrong decisions can be costly. In this way, some of the decision-making problems that a manager commonly faces are, for example, the decision to determine a price, the decision to buy or manufacture, in logistics problems, storage problems, etc. On the other hand, data collection has become a competitive advantage, as it can be used for the analysis and extraction of significant results throughtheapplicationofvarioustechniques,suchasstatistics,simulation,mathematics,econometrics and current techniques, , such as machine learning to create predictive models. In addition, there is evidence in the literature that the creation of models with machine learning techniques has a positive impact on the industry and in different areas of research. In this context, the present work proposes the development of a predictive model for decision making, using supervised machine learning techniques, and combining the model generated with the constraints pertaining to the optimization process. The purpose of the proposal is to train a mathematical model with historical data from a decision-making process and obtain predictors composed of empirical functions that will later be used and modeled according to the constraints of the problem. Thus, this work can be classified as applied research, with descriptive empirical objectives and practical experience that will explain the model and its advantages. The way to approach the problem of this work will be quantitative, being the technical procedures of modeling and simulation. The proposed system is validated by applying it to a real problem in a Brazilian multinational company in the auto parts segment, located in Vale do Paraíba-SP. The data set comes from the transport process, aiming to apply the systematic and create a mathematical model to predict the cost of transporting a quantity of items to a specific destination, so that this model finally helps to reduce the estimation times and take better decisions. Finally, for the creation of the predictive model, the supervised learning techniques of linear regression, CART decision tree, M5P model tree, support vector machine and neural networks were tested. All the models generated were evaluated by the statistical metrics of performance of the mean absolute error, root of the mean square error, relative absolute error, relative quadratic error, correlation coefficient and determination coefficient or also known as R2. Since the M5P model obtained the lowest values in the error metrics and the highest values in the correlation and R2, showing to be an efficient model, in addition to providing the necessary equations for the optimization process
In different aspects of everyday life, the human being is forced to choose between several options, this process is known as decision making. At the business level, decision making plays a very important role, because the success or failure of organizations depends on these decisions. However, in many cases, making the wrong decisions can be costly. In this way, some of the decision-making problems that a manager commonly faces are, for example, the decision to determine a price, the decision to buy or manufacture, in logistics problems, storage problems, etc. On the other hand, data collection has become a competitive advantage, as it can be used for the analysis and extraction of significant results throughtheapplicationofvarioustechniques,suchasstatistics,simulation,mathematics,econometrics and current techniques, , such as machine learning to create predictive models. In addition, there is evidence in the literature that the creation of models with machine learning techniques has a positive impact on the industry and in different areas of research. In this context, the present work proposes the development of a predictive model for decision making, using supervised machine learning techniques, and combining the model generated with the constraints pertaining to the optimization process. The purpose of the proposal is to train a mathematical model with historical data from a decision-making process and obtain predictors composed of empirical functions that will later be used and modeled according to the constraints of the problem. Thus, this work can be classified as applied research, with descriptive empirical objectives and practical experience that will explain the model and its advantages. The way to approach the problem of this work will be quantitative, being the technical procedures of modeling and simulation. The proposed system is validated by applying it to a real problem in a Brazilian multinational company in the auto parts segment, located in Vale do Paraíba-SP. The data set comes from the transport process, aiming to apply the systematic and create a mathematical model to predict the cost of transporting a quantity of items to a specific destination, so that this model finally helps to reduce the estimation times and take better decisions. Finally, for the creation of the predictive model, the supervised learning techniques of linear regression, CART decision tree, M5P model tree, support vector machine and neural networks were tested. All the models generated were evaluated by the statistical metrics of performance of the mean absolute error, root of the mean square error, relative absolute error, relative quadratic error, correlation coefficient and determination coefficient or also known as R2. Since the M5P model obtained the lowest values in the error metrics and the highest values in the correlation and R2, showing to be an efficient model, in addition to providing the necessary equations for the optimization process
Descrição
Palavras-chave
Aprendizagem de Máquina, Otimização, Tomada de decisão, Árvore de decisão, Algoritmo M5P, Machine Learning, Optimization, Decision making, Decision tree, M5P algorithm, Processo decisório, Otimização matemática, Algoritmos de computador