A connectionist approach for building influence diagrams

Abstract

The development of adaptive systems must face the problem of recognition as a synergy of learning and knowledge. This paper presents a method for constructing influence diagrams from backpropagation neural networks, as a way of combining the main advantages of these methodologies. The basic concepts of influence diagrams and neural networks are discussed as a brief review. An algorithm to extract the conditional probabilities of the network is presented and illustrated by three pattern recognition examples. Although much of the a priori information from the sample set is lost during the training phase of the network, an influence diagram that behaves as the original knowledge source can be constructed.