Definição de parâmetros de RBF utilizando grafo de Gabriel

Abstract

The use of RBF (Radial Basis Function) in sample classication problems has been a much discussed approach in the literature. Its diverse applications allow you to deal with problems of dierent complexities. Among the most applied radial functions Gaussian is considered one of the most ecient for its simplicity of conguration. However one of the challenges in using this type of function is to denethe parameters c and sigma, center and radius respectively, suitableto avoid sub or oversize in the solution of the problem. This workproposes a methodology based on Gabriel's graph and the theory ofthe dominant sets to nd these parameters for a given set of samples,without the necessity of arbitration. In a second step, the gaussiansfound in an articial neural network architecture are applied in orderto classify these samples. In a third step, we compare the resultsfound with those of classical classiers known in the literature. Oncethese results are analyzed, the particularities of each problem studiedand their inuences on the metrics of the proposed methodology areanalyzed, which may create a need to adapt them to the treatment ofsome of these particularities. Among the particularities studied arethe overlap of classied samples, dened as noise, and the unbalanceof samples, very common in real problems. In order to deal with theoverlap a ltering process was proposed that aims to improve the accuracy in the classication. For the unbalance the technique was usedin the technique of undersampling which seeks to improve accuracyas described in the literature. In order to re-embroider a problem ofclassication of dispatch of plants, it was proposed to use the methodin a six-bar system with two thermoelectric plants, one hydroelectricplant and two wind farms classied as distributed generation, to validateIts application in a problem that is very current, ie, which typeof plant should be dispatched depending on the climatic conditions.Finally, the proposed method aims to treat the described points in orderto obtain the best possible results for each problem, without theneed to adjust Gaussian parameters a priori, which allows its application, in general, to be simple to congure, not needing deep technical knowledge for its use.