Uma nova abordagem baseada em margem para seleção de modelos neurais

Abstract

This work presents a new decision-making strategy to the multiobjective learning of artificial neural networks. The objective is to find the solution within the pareto-optimal set that has the best generalization performance. The proposed decision-making approach is based on a geometric approximation to the maximum margin (distance) of class separation, which is estimated through the following steps: modeling of input patterns using the gabriel graph, detection of class separation borders and synthesis of patterns along the maximum margin region. This methodology allows the selection of smooth (that ignore noise) and well-fitting models in a straightforward manner, i.e., without the need of the tuning of parameters by the user or the use of a representative validation data set. Results on benchmarks in literature showed that our decision-making method, combined with multiobjective training, was efficient to control the generalization of neural models.