Structural Learning of Neural Network for Continuous Valued Output: Effect of Penalty Term to Hidden Units

Abstract

Multilayer feed forward networks with back propagationlearning are widely used for function approximation but the learned networks rarely reveal the input output relationship explicitly. Structural learning methods are proposed to optimize the network topology as well as to add interpretation to its internal behaviour. Effective structural learning approaches for optimization and internal interpretation of the neural networks like structural learning with forgetting (SLF) or fast integration learning (FIL) have been proved useful for problems with binary outputs. In this work a new structural learning method based on modification of SLF and FIL has been proposed for problems with continuous valued outputs. The effectiveness of the proposed learning method has been demonstrated by simulation experiments with continuous valued functions.