A Meta Neural Network Polling System for the RPROP Learning Rule

Abstract

This paper proposes an application independent method of automating learning rule parameter selection using a group of supervisor neural networks, known as meta neural networks, to alter the value of a learning rule parameter during training. Each meta neural network is trained using data generated by observing the training of a neural network and recording the effects of the selection of various parameter values. A group of meta neural networks is then polled to obtain a parameter value for a learning rule. Experiments are undertaken to see how this method performs by using it to adapt a global parameter of RPROP.