Research Seminar - March 26, 1999

Abstract

Genetic algorithms (GAs) are a well known optimisation technique that has been used extensively during the past 20 years. The standard GA will create a population of strings that are modified by the application of selection, crossover, and mutation operators. The effectiveness of GAs is based on their ability to parse fit substrings at higher rates than unfit substrings. This string parsing ability is the embodiment of Holland's Schema Theorem. However, the Schema Theorem requires that the substrings that are being parsed are meaningful "building blocks" that can be reconstructed into whole strings that are also fit. This is the Building Block Hypothesis.

Neural Networks (NNs) are a distributed computing structure. The neurons are small units which apply a function over the sum of their inputs, and then transmit this value to all outputs from the neuron. Neurons are connected by weighted connections, which can modulate the signal received and transmitted from each connection. Traditionally, to evoke a particular behaviour from a NN requires training, which is most often performed by modifying the values of the weights on the connections between neurons.

A relatively new field is the use of GAs to evolve weights of connections for NNs. This approach has been useful in evolving NNs which solve a variety of problems that were previously considered "difficult". However, the current understanding of how this evolutionary technique works is limited to very simple problems and NNs.

In this talk, I will describe how different methods of representing NNs to a GA alter the results obtained from the GA. Additionally, I will explain why it is possible for GAs to evolve weights for a distributed architecture, and why the Building Block Hypothesis holds in these instances.