| Resumo: | In optimization problems there is a subset of problems that are defined as complex problems, which present high complexity models. For this class of problems there is an exhaustive number of possible combinations for the input variables of a system. Thus, evaluating these combinations is a humanly unfeasible process, so we use optimization mechanisms that aim to find the best solution, among which it is possible to quantify the degree of adequacy of the solutions to the needs in question. Generally, when dealing with problems with up to three objective functions, Evolutionary Algorithms are used to solve them. Another approach employed is the use of surrogates, which can be defined as mechanisms capable of learning the behavior of a given function. When using these mechanisms in complex problems, it is estimated that the high computational cost reduction to obtain the fitness values of the objective functions will be gained. Among the common surrogate mechanisms in the literature, the techniques of linear regression and machine learning stand out. The application of surrogates in problems with more than one objective function, multiobjective problems, requires the use of a learning model for each function, however, recent studies have been successful in employing a single surrogate for problems with more than one objective function. However, the use of surrogate in optimization problems with more than three objective functions is still a little explored area. Therefore, this work aims to propose and evaluate new approaches to surrogate training associated with Evolutionary Algorithms. Two frameworks were developed, one applied to the class of mono-objective problems and the other aimed at optimization problems with many objectives.The proposed frameworks are characterized by the use of different approaches to surrogate training and also different ways of using machine learning techniques. The frameworks were subjected to experiments using benchmark problems, where each configuration of the algorithms was executed for twenty times and stores the performance metrics. To confirm or refute the hypotheses, the Wilcoxon statistical test was applied. The results show that the machine learning techniques, Decision Tree and Random Forest when applied as a surrogate provide satisfactory results, in addition, the surrogate training methodologies proposed here, associated with the NSGA-II and SMPSO algorithms obtained better or equal results. than state-of-the-art algorithms (NSGA-II and MOEADD), in most of the experiments carried out.
|
|---|