RBF Network under imperfect situation

Abstract

Although many methods are existed to handle multiple kinds of failures in radial basis function (RBF) network training, there is no effective and efficient method to select RBF centers. We add a 𝑙1 norm sparsity regularization term into the original failure tolerant training objective function. Since the 𝑙1 norm regularization term has an ability to turn some unnecessary RBF weights to zero, the trained network based on the modified objective function will contain essential RBF nodes only. Since the modified objective function is non-differentiable, traditional optimization method cannot be used to minimize the modified objective function. In this project, I investigate an analog method, from the concept of the local competition algorithm (LCA), to handle the non-differentiable objective function. Simulation result conveys that my method is better than the training method of orthogonal least square (OLS).