Building expert systems by training with automatic neural network generating ability

The authors examine the construction of a connectionist expert system without specifying the network structure before training. The generated connectionist expert system consists of many features, such as operation of forward and backward inference based on partial input information, online learning, noisy data handling, generalization, and the explanation ability. Two sample problems, the Knowledge Base Evaluator 1 and Treatment of Posiboost, are considered in order to illustrate the workings of the connectionist expert system. The training algorithm, which has network generating ability, is presented to build the knowledge base of the connectionist expert system. It provides the abilities needed to realize the described features of the connectionist expert system. This proposed system can be easily used to build expert systems quickly, and the inferencing in the developed systems will be fast