The Effects of Problem Representation and Network Representation on Training Results of Artificial Neural Networks

Abstract

There are different ways to obtain a good Artificial Neural Network. When training, the choice of the data set is of importance to the quality of the resulting network. When evolving a network using Genetic Algorithms, it is important that the representation of the network does not interfere with the passing-on of information to next generations. I looked into the effects of data representation on the quality of the trained networks, and I investigated one solution proposed by Thierens (1996) to unheuristically remove redundancies in genotype. I could not verify the results found in the proposed solution.