Simulation of a Multi-Dimensional Pattern Classifier

PhD Thesis, 1996

Andy Cheetham

Abstract

Current techniques for multi-dimensional pattern recognition are examined with particular emphasis on the use of artificial neural networks. A solution in the form of a Self-Organising and Self-Adaptive (SOSA) network algorithm is devised and simulated to offer a new architecture and training methodology. This network greatly reduces training times while preserving the relationships among input elements. Furthermore, the SOSA network offers the advantage of becoming simplified as training progresses. The implications of the unique properties of the SOSA network are presented.

To verify the quality of the proposed SOSA network, extensive simulation results are obtained and presented. The SOSA network is applied to a number of practical examples including some that involve 3-dimensional surface recognition. The effect of varying the network parameters and their impact on the network properties is explored. The susceptibility of the network to noisy and incomplete data, as well as its ability to recognise shifted and rotated patterns, is investigated. Other examples presented include printed character and acoustic emission patterns, the latter being an application area represented by a 1-dimensional input space.

Conclusions are drawn as to the suitability of the SOSA network to multi-dimensional pattern recognition, and a plan for future work is proposed.