Integration of Neural Heuristics into Knowledge-based Inference

Abstract The rule base and the inference engine of a knowledge-based system are transformed into a kind of neural network called a conceptualization network. An approach is presented that generalizes the backpropagation teaming rule of the neural-network approach such that it can effectively deal with errors in conceptualization networks, which are often multilayered and involve logic conjunction. The idea is to use hill-climbing search where the backpropagation rule falls short because the transfer function is not differentiable. When the generalized backpropagation rule is applied to a conceptualization network which has been constrained by initial correct knowledge, incorrect rules can be recognized. Experiments in a practical domain have demonstrated that the approach can satisfactorily conduct credit and blame assignment for rules which may involve intermediate concepts and logic conjunction. Effective removal of incorrect rules with significant improvement of the system performance has been observed. KEYWORDS: Knowledge-based systemsneural networksbackpropagationcredit and blame assignmentconceptualization networks.