Connectionist expert systems

Two realizations of connectionist expert systems (shells) which facilitate building expert systems when raw data and/or expert rules are available are presented. The knowledge base is represented as a neural network trained either by using past data or using rules. The systems facilitate approximate reasoning, creating a user interface or a communication with an object in real time, explanation to the user, learning and adaptation of the existing knowledge during the working phase, learning explicit rules about the domain area, and learning fuzzy rules in particular. The two different environments reported depend on the standard neural network simulators used. These two environments have been experimentally used for creating two diagnostic expert systems: one for breast-cancer diagnosis, another for fault diagnosis of an electronic device.<<ETX>>