Resumo:
In the context of industry 4.0, optimization via simulation (OvS) emerges as one of the most powerful tools in the modern industry, allowing decision-makers to allocate their resources more assertively. However, in very complex systems, the use of conventional OvS techniques requires computational time, which frequently, makes its application unfeasible. In recent years, the development in the machine learning area has emerged algorithms with high learning capacity, making the use of optimization via simulation by metamodeling (OvSM) techniques to solve complex problems a promising field of study. In this sense, the present study proposes a framework for OvSM based on the insights and analyses derived from the systematic literature review carried out. The proposed framework incorporates the use of discrete event simulation techniques, design of experiments, machine learning algorithms, and hyper-parameter optimization via genetic algorithm for OvS problems. To validate the proposed method, this dissertation tested and compared six machine learning algorithms (Support Vector Machine, Artificial Neural Networks, Gradient-Boosted Trees, Randon Forest, Polynomial Regression, and Gaussian Process) with and without the hyper optimization step -parameters in two experimental arrangements (Latin Hypercube Design and Random) applied to the problem of resource allocation in three real cases in the industry. With the application of the method in the study objects presented, the best performing metamodels obtained solutions that reached, respectively, 100%, 96.17%, and 100% of the optimal benchmark location, demanding, on average, 35.22% less time computational. Also, the incorporation of the hyper-parameter optimization step in the proposed metamodeling method allowed a 31.28% reduction in the root mean square error of the metamodels compared to the traditional method, which does not include this step.