Experience-based learning in deductive reasoning systems

General knowledge is widely applicable, but relatively slow to apply to any particular situation. Specific knowledge can be used rapidly where it applies, but is only narrowly applicable. We present an automatic scheme to migrate general knowledge to specific knowledge during reasoning. This scheme relies on a nested rule representation which retains the rule builder's intentions about which of the possible specializations of the rule will be most useful. If both general and specific knowledge is available and applicable, a system may be slowed down by trying to use the general knowledge as well as, or instead of, the specific knowledge. However, if general knowledge is purged from the system after migration, the system will lose the flexibility of being able to handle different situations. To retain the flexibility without paying the price in speed, a shadowing scheme is presented that prevents general knowledge from being used when specific knowledge migrated from it is available and applicable. The combination of knowledge migration and knowledge shadowing allows a deductive reasoning system to learn from and exploit previous experience. Experience is represented by the instance relationship between the general knowledge and the specific knowledge migrated from it. We also present techniques for implementing efficient rules of inference in natural deduction systems by caching and recalling the history of rule activation steps that alleviate duplicate pattern matchings and binding conflict resolutions. To reduce the complexity of manipulating rule activation steps from exponential to polynomial, methods of distributing the information about rule activation steps are developed that minimize the number of activation steps and the number of substitution compatibility tests among shared variables. An implementation of these schemes in a network-based reasoning system is discussed. Test results are shown that demonstrate the predicted benefits of these ideas.