A MAXIMUM LIKELIHOOD APPROACH FOR RBF CENTER SELECTION IN CLASSIFICATION PROBLEMS

This paper proposes a binary classification method that transfers the geometric interpretation of support edges to RBF (Radial Basis Function) networks. Starting from the geometric motivation provided by Chipclass and Gabriel Graphs (GG), we describe the evolution of G-RBF toward G-RBF-S, a probabilistic formulation grounded in Maximum Likelihood Estimation (MLE) and Poisson processes. In synthetic experiments, the method outperforms the exhaustive-search approach (Full Gabriel) while achieving speedups up to 267.4x. Across 15 real datasets, tuned G-RBF-S attains a better mean rank than the exhaustive model and remains statistically equivalent to it.