Resumo:
A pillar of Industry 4.0, Simulation Optimization (SO) is a powerful tool used across several fields, which allows the evaluation of a system under different conditions, facilitates its performance analysis and makes decision-making more efficient. On the other hand, the SO might be time-consuming, particularly when considering complex model optimization. In this sense, metamodeling has emerged as a promising technique for simulation optimization. Metamodeling aims to establish and estimate a relationship between the inputs and outputs of a simulation model, creating a simplified model used to evaluate potential solutions during the optimization process. Metamodeling approaches can be classified as metamodeling with a fixed experimental design, in which a single experimental database is generated at the beginning of the project and the metamodel is trained exclusively on this basis, and adaptive metamodeling, which is based on the iterative construction of metamodels that are adjust and refine throughout the optimization process. This paper presents a novel metamodeling framework, called Adaptive Metamodeling-based Simulation Optimization (AMSO), to optimize complex and expansive discrete event simulation models. The proposed approach combines machine learning and metaheuristic techniques to identify the most promising areas of the solution space that can be explored more efficiently to achieve high-quality solutions. The proposed framework is evaluated in three real-world industry case studies: a resource allocation problem in a manufacturing digital twin model, a capacity expansion project for an iron ore plant, and an allocation problem in a temporary hospital. Compared to the Efficient Global Optimization method, AMSO found a solution that was 8.1% better (on average) in the first case study, 9.7% better in the second, and 28.4% better in the third, with no significant difference in computational time spent. Additionally, AMSO found solutions statistically equivalent to the Genetic Algorithm method but required 83.6%, 90.6%, and 65.2% less computational time in the first, second, and third cases, respectively. The results found demonstrate the applicability and robustness of the proposed framework for optimizing the study objects analyzed