A novel neural network approach to knowledge acquisition
Mu‐Chun Su · 1993
Often a major difficulty in the design of rule-based expert systems is the process of acquiring the requisite knowledge in the form of production rules. This dissertation presents several classes of composite neural networks which are trained to provide appealing solutions to the problem of knowledge acquisition. The value of the network parameters, after sufficient training, are then utilized to generate production rules on the basis of preselected meaningful coordinates. Neural networks are attracting a lot of interest in the scientific community because of their dynamical nature: robustness, capability of generalization and fault tolerance. Neural networks have already proven useful in low level information processing (e.g. signal analysis). However, an apparent disadvantage of traditional neural networks (i.e. backpropagation networks) is that they do not provide explanations of their response i.e. explicit rules or logical reasoning processes. The classes of composite neural networks studied in this dissertation integrate the paradigm of neural networks with the rule-based approach, rendering them more useful than either. It is shown that the composite neural networks act as Bayesian classifiers, and more importantly, can provide density estimates of the class variables. The dissertation provides a mathematical framework for achieving reasonable generalization properties of composite neural networks serving as universal approximators via an appropriate training algorithm (supervised decision-directed learning). The algorithm is based on an approach that divides the input (feature) space into proper subsets (e.g. hyperspheres, hyperrectangles, or hyperellipsoids), represented as portions of the trained composite neural networks, while the complete structures provide acceptable knowledge representations of the data. The new concepts and methods presented in the dissertation are illustrated via two traditional academic pattern recognition examples and one practical example from medical diagnosis. These examples are discussed in detail. Finally the dissertation concludes with a summary of the most important results accomplished and with brief recommendation for future work.