Estratégias de decisão em aprendizado de máquina multi-objetivo

Abstract

This work addresses the problem of selection models obtained through multi-objective machine learning and presents strategies for choosing the Pareto-optimal set. Within the multiobjective approach, several options of candidate solutions are obtained, characterized by non-dominance among themselves. As part of the multi-objective approach, a decision procedure is needed among the set of candidate solutions that have been generated. In machine learning, the decision criterion should portray the dilemma of the balance between polarization and variance effects, the central theme of machine learning, and indicate a solution that best represents this balance. The decision strategies proposed in this work were defined for two of the main problems of supervised learning: classification and regression. These strategies are characterized by being independent of resampling and the structure of the model used. The ability to use new information to help in the model selection process has ensured advances in addressing the dilemma between polarization and variance in machine learning. The numerical results, through learning problems with artificial and real data, were evaluated with other known decision strategies, such as: the Curve L method, the validation error method and, in addition, compared with the results presented by other learning algorithms such as the Support Vector Machines. With the results presented, the decision methods allowed the training algorithms to have a better use of the original dataset and, consequently, an improvement in the generalization capacity. Thus, the decision process in supervised machine learning, from the perspective of multiobjective optimization, brought a new model selection script according to the problem in question, making the procedure well-structured and deterministic.