Exploring the Relationship between Neural Network Topology and Optimal Training Set by Means of Genetic Algorithms

Abstract

In a previous paper I have presented the results of optimizing Neural Network (NN) topology for the task of Natural Language Processing (NLP). In that research, all NN were trained with a fixed 20% of the total language. In this paper I present results of optimizing a set of configuration values that have proven to affect NN performance. For example, Elman has reported improved performance when the NN were trained with simple sentences first, and complex sentences later. On the other hand, Lawrence, Giles, and Fong have reported better results when the training data was presented in a single set. Lawrence, Giles, and Fong have also studied the effect of different learning algorithms on natural language tasks. Because of the ability of GA to search a problem space for minima without using knowledge about the problem itself, they are well suited for problems that might contain more than one possible solution. Finding different minima becomes important for real-life applications, since variables such as number of hidden nodes, number of hidden layers, number of connections, and size of training set all can affect training and response time for NN.