Re nement of Approximate Domain Theories byKnowledge-Based Neural Networks

Standard algorithms for explanation-based learning require complete and correct knowledge bases. The KBANN system relaxes this constraint through the use of empirical learning methods to reene approximately correct knowledge. This knowledge is used to determine the structure of an artiicial neural network and the weights on its links, thereby making the knowledge accessible for modiication by neural learning. KBANN is evaluated by empirical tests in the domain of molecular biology. Networks created by KBANN are shown to be superior, in terms of their ability to correctly classify unseen examples, to randomly initialized neural networks, decision trees, \nearest neighbor" matching, and standard techniques reported in the biological literature. In addition, KBANN's networks improve the initial knowledge in biologically interesting ways.