KNOWLEDGE-BASED CONNECTIONISM FOR REVISING DOMAIN THEORIES

Integration of domain theory into empirical learning is important in building a useful learning system in practical domains since the theory is not always perfect and the data is not always adequate. A novel knowledge-based connectionist model referred to as KBCNN for machine learning is presented. In the KBCNN learning model, useful domain attributes and concepts are first identified and linked in a way consistent with initial domain knowledge, and then the links are weighted properly so as to maintain the semantics. Hidden units and additional connections may be introduced into this initial connectionist structure as appropriate. Then, this primitive structure evolves to minimize empirical error. The KBCNN learning model allows the theory learned or revised to be translated into the symbolic rule-based language that describes the initial theory. Thus, a domain theory can be pushed onto the network, revised empirically over time, and decoded in symbolic form. The domain of molecular genetics has been used to demonstrate the validity of the KBCNN learning model and its superiority over related learning methods.