A probabilistic rule-based system in artificial neural networks

The authors propose a hybrid system in which a probabilistic rule-based system is implemented as a forward chaining inference machine. Based upon this design, the rule-based system could be equipped with a generalization function, automatic rule learning functions and a damage tolerant feature. The system contains three networks, Hopfield binary neural network, a single-layered feedforward neural network and a multi-layered feedforward neural network, and operates in two distinctive phases for the network training and inference operations. In the training phase, a set of common knowledge pieces as information units and the inference rules used to derive new information units are both implemented in the system in order to assign a certain problem domain to the system. In the inference phase, the system accepts information units from the user and infers new information units, according to the common knowledge and the rules