Optimisation of a weightless neural network using particle swarms

Among numerous pattern recognition methods the neural network approach has been the subject of much research due to its ability to learn from a given collection of representative examples. This thesis is concerned with the design of weightless neural networks, which decompose a given pattern into several sets of n points, termed n-tuples. Considerable research has shown that by optimising the input connection mapping of such n-tuple networks classification performance can be improved significantly. In this thesis the application of a population-based stochastic optimisation technique, known as Particle Swarm Optimisation (PSO), to the optimisation of the connectivity pattern of such “n-tuple” classifiers is explored.

The research was aimed at improving the discriminating power of the classifier in recognising handwritten characters by exploiting more efficient learning strategies. The proposed "learning" scheme searches for ‘good’ input connections of the n-tuples in the solution space and shrinks the search area step by step. It refines its search by attracting the particles to positions with good solutions in an iterative manner. Every iteration the performance or fitness of each input connection is evaluated, so a reward and punishment based fitness function was modelled for the task. The original PSO was refined by combining it with other bio-inspired approaches like Self-Organized Criticality and Nearest Neighbour Interactions. The hybrid algorithms were adapted for the n-tuple system and the performance was measured in selecting better connectivity patterns. The Genetic Algorithm (GA) has been shown to be accomplishing the same goals as the PSO, so the performances and convergence properties of the GA were compared against the PSO to optimise input connections.

Experiments were conducted to evaluate the proposed methods by applying the trained classifiers to recognise handprinted digits from a widely used database. Results revealed the superiority of the particle swarm optimised training for the n-tuples over other algorithms including the GA. Low particle velocity in PSO was favourable for exploring more areas in the solution space and resulted in better recognition rates. Use of hybridisation was helpful and one of the versions of the hybrid PSO was found to be the best performing algorithm in finding the optimum set of input maps for the n-tuple network.