We have viewed incremental learning in terms of search and belief revision. Our analysis has been cursory given the simplistic nature of knowledge bases constructed by most inductive programs. While systems like COBWEB, UNIMEM, CYRUS, and PROTOS are introducing organizations of greater complexity, the important characteristics and tradeoffs associated with them are not well understood. Paralleling our empirical studies, we will flesh out the relationship between inductive learning and belief revision techniques. Undoubtedly, there will be advantages for each field.
Belief revision techniques can suggest how to scale up inductive learning methods. For example, while STAGGER and a few other systems can track environmental drift, they do so by `destroying' previously-learned conceptualizations. However, this strategy can undo useful information - we do not want to have to relearn important lessons about driving as we cycle between Summer and Winter. To our knowledge, no incremental learning system can track change without destroying prior knowledge.
Research into the problem of tracking concept drift will look for alternates to the justification-based belief revision techniques (Doyle, 1979) in which our discussions have been implicitly framed. In particular, DeKleer's (1986) assumption-based truth maintenance system (ATMS) allows the simultaneous existence of multiple `world views' and suggests a viable framework in which to explore sophisticated tracking mechanisms. The abilities of STAGGER and ID4S will be extended to partition knowledge into distinct and internally-consistent conceptual blocks. Tracking will also be extended to COBWEB, which must simultaneously support multiple prediction tasks.
Finally, our work on inductive learning assumes that perfect consistency may not be attainable. In theory, truth maintenance systems may be able to deal with noisy domains, but not in an efficient manner. Knowledge base organizations that exploit probabilistic representations and conservative revision strategies may lead to parsimonious and efficient belief revision methods. A unification of induction and belief revision mechanisms will also shed light on the continuum between data-driven and knowledge-driven (e.g., explanation-based) learning methods. Work on these issues will undoubtedly be well underway by the end of the second year, but will most likely not reach fruition until afterwards. Thus, they will be a natural basis for extensions to this proposal.