GLOBAL OPTIMIZATION OF NEURAL NETWORKS AND BEHAVIOUR EXPLANATION

Abstract

In this thesis we study two aspects of neural networks in practical applications: optimization, and behaviour explanation. Firstly, we examine how genetic algorithms are used to globally optimize a neural network by [1] selecting relevant input attributes, [2] determining optimal network architecture, and [3] optimizing connection weights. Experimental results indicate an improvement in prediction accuracy, and significant reduction in training and prediction time. However, a major drawback is the amount of time required for evolution to take place. Secondly, we investigate a set of paradigms to analyze the behaviour of a trained neural network. The analysis enables us to provide partial explanations for the decision made by the neural network, and helps to dispel some of the anxiety involved with the perception of neural networks as "black boxes".