BACKPROPAGATION NEURAL NETWORKS FOR FUZZY-LOGIC

Expert systems play a crucial role in the application of artificial intelligence to real problems. There are numerous methods of dealing with uncertainty in expert systems. Fuzzy logic is one such approach, in which the uncertainty is represented by possibility distributions for the antecedents and consequent of a rule. An uncertain input can be propagated to the consequent by a variety of mechanisms. This generality results in a higher computational burden for the system. More importantly, a means of controlling the response of the system to variations in input conditions is needed. In this paper, feedforward neural networks are proposed as a means of controlling this generality as well as providing the parallel computation necessary for fuzzy logic. It is demonstrated that these networks can learn and extrapolate complex relationships between antecedents and consequents for rules containing single and conjunctive antecedent clauses. Also, multiple compatible rules can be stored in a single network structure, increasing the efficiency of rule storage, and providing a natural mechanism for conflict resolution.