Using neural nets to learn weights of rules for compositional expert systems

Knowledge base for a compositional expert system consists of a set of IF THEN rules with uncertainties expressed as weights. During consultation for a particular case, all aplicable rules are combined and weighhs of goals (final diagnoses or recommendations) are computed. The main problem when eliciting such knowledge base from an expert is the question of “correct” weights of rules.Our idea is, to combine the structure of knowledge obtained from expert with weights learned from data. We choose the topology and initial settings of the neural net (number of neurons, prohibited links) according to the rules obtained from expert. Then, after learning such network, we try to interpret the weights of connections as uncertainty of the original rules.The paper shows some experimental results of this approach on a knowledge base for credit risk assessment.