Extracting knowledge from a hybrid, symbolic connectionist network
Steve G. Romaniuk · 1992
This dissertation describes the basic features of the hybrid symbolic, connectionist learning system SC-net. The main objective of SC-net is to overcome the knowledge acquisition bottle-neck in the development of expert systems. It uses the recruitment of cells learning algorithm for the construction of a network topology. The global attribute covering algorithm (GAC) can then be invoked to find sub-optimal minimal covers of the example space. Because of its hybrid structure knowledge can then (after learning) be extracted in the form of rules. Uncertainty management both during learning and testing has been included. A modification of the GAC algorithm has been developed to allow learning of fuzzy pi-shaped membership functions, when used in conjunction with variables. Several natural domains and well known learning algorithms have been selected for comparison purposes. Finally, an investigation of the time and space complexity of the SC-net model has been provided. The analysis is tailored towards a software and a hypothetical connectionist hardware model.