The proposed research is aimed at developing a systematic artificial neural network (ANN) design and training methodology and exploring its applications in robotics. The salient feature of the research is the use of statistical pattern recognition techniques to solve some key aspects of ANN and learning procedures. One major limitation of present ANNs operating under the supervised mode of learning is that these networks are unable to self-configure the architecture for a given classification problem. A solution to this problem is to establish a direct relationship between a class of nonparametric classifiers, i.e. decision trees, and the multilayer neural networks. It can be shown that a decision tree can be restructured as a three layered neural net. Exploiting this restructuring allows neural network design and training methodology to have self-configuration capability. The tree-to-network mapping also provides a solution to the credit assignment problem thus making it possible to train each layer separately and progressively. The research will explore automatic tree generation, incorporation of incremental learning, and adaptability in such networks. The proposed methodology will be applied to robot learning where networks are required to perform regression rather than classification. The main expected benefit of the proposed research is that it will make available an ANN design and training procedure that is systematic. This should result in many more applications for artificial neural networks.
