Método para calibração de parâmetros com abordagem adaptativa de busca por agrupamento e modelos estocásticos

• Ano de defesa: 2021
• Autor(a) principal: Nascimento, Marcelo Branco do [UNIFESP]
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Paulo
• Palavras-chave em Português: Métodos Heurísticos; Calibração de Parâmetros; Redes Bayesianas; Algoritmo de Estimação de Distribuição; Métodos Estocásticos
• Link de acesso: https://hdl.handle.net/11600/62515

Finding a solution to an optimization problem involves looking for good results in a solution space. This task is computationally very costly, as it handles efficiently and robustly. The use of meta-heuristics falls into the category of methods used to solve these problems. Therefore, an appropriate adjustment of the metaheuristic parameters becomes important. This task is considered difficult, as there are differences between the nature of the parameters. To solve this tuning problem, this research developed the Bayesian Network Tuning (BNT). A method to tuning meta-heuristic parameters, using stochastic and adaptive strategies, which use statistical information of dependence between different parameters with a cluster search technique. Thus, it is intended to develop an efficient parameter tuning method based on the methodology of Distribution Estimation Algorithms, to assist meta-heuristics in solving optimization problems. Two validations were performed. The first, using AClib to compare with other tuning methods in different scenarios. In the second validation, a case study of the proposed method was carried out on the meta-heuristic Biased Random-Key Genetic Algorithm (BRKGA) to solve the 1D Packaging Problem (BPP 1D). In both results, BNT was competitive and robust compared to other tuning methods found in the literature.