RBF networks with concurrent fault situation

Abstract

As a neural network model, RBF network has been extensively studied in these recent two eras. At the present stage, the most effective way to handle node and weight failure situation, which happens a lot in RBF network, is yet to be discussed. General speaking, there are four kinds of node and weight failure in RBF networks. They are open node fault, open weight fault, node noise and weight noise. Studies have been done generally to deal with situation where only one kind of failure is in a network. Last year, my fellow did a project to handle four concurrent node fault situation using orthogonal least square (OLS) method. However, it requires an additional learning algorithm for selecting the RBF centers. In this project, we add a 𝑙1 norm regularizer into the fault tolerant objective function. During training, it can automatically eliminate unimportant RBF nodes. In this approach, we need to the weighting factor to control the number of the selected. To minimize the objective functions, I use the alternating direction method of multipliers (ADMM) framework. ADMM is an algorithm that solves convex optimization problems. In my approach, no extra algorithm is required to select the RBF centres. Simulation results show that my approach is superior than the OLS based approach.