Abstract:
Research efforts in the improvement of artificial neural networks have provided significant enhancements in learning ability, either through manual improvement by researchers or through automated design by other artificial intelligence techniques, and largely focusing on the architecture of the neural networks or the weight update equations used to optimize these architectures. However, a promising unexplored area involves extending the traditional definition of neural networks to allow a single neural network model to consist of multiple architectures, where one is a primary architecture and the others supplementary architectures. In order to use the information from all these architectures to possibly improve learning, weight update equations are customized per set-of-weights, and can each use the error of either the primary architecture or a supplementary architecture to update the values of that set-of-weights, with some necessary constraints to ensure valid updates. This concept was implemented and investigated. Grammatical evolution was used to make the complex architecture choices for each weight update equation, which succeeded in finding optimal choice combinations for classification and regression benchmark datasets, the KDD Cup 1999 intrusion detection dataset, and the UCLA graduate admission dataset. These optimal combinations were compared to traditional single-architecture neural networks, which they reliably outperformed at high confidence levels across all datasets. These optimal combinations were analysed using data mining tools, and this identified clear patterns, with the theoretical explanation provided as to how these patterns may be linked to optimality. The optimal combinations were shown to be competitive with state-of-the-art techniques on the same datasets.
