next up previous contents
Next: Frameworks for understanding incremental Up: Plan for Continued Research Previous: Domain-independent incremental methods

Application areas for incremental learning

As we proceed, we will test our general learning methods in a number of application areas. For example, STAGGER's ability to incrementally introduce new terms through chunking is important for adaptive problem solving. Quinlan (1983) recognized the need for this capability in learning to recognize chess-end games. We plan to highlight and extend STAGGER's ability in the context of Samuel's (1967) Checker's program.

Augmenting the Concept Language During Game-Playing (Man-Machine Studies, 1989)

Samuel's original Checker's program learned by adjusting the weights of a linear equation of primitive board features. Later versions overcame the fundamental weaknesses of this approach by using a signature table, which explicitly `hard-wired' feature combinations. This paper will report on STAGGER's ability to incrementally add new terms (i.e., chunks) to the feature description language, thus overcoming the need to manually derive these feature combinations.

Intelligent tutoring (Wenger, 1987) is another area in which we will test our systems. While an effective tutor must adapt along a number of fronts, we are currently focusing on the problem of incremental student modeling.

Incremental Student Modeling (Man-Machine Studies, 1990)

Langley, Ohlsson, and Sage (1984) used ID3 to construct a student model from buggy behavior. Their system accounted for many of the subtraction `bugs' cataloged by VanLehn (1982). Unfortunately, ID3 is nonincremental, and as observations were accumulated lag times became significant. We plan to incorporate ID4S into the student modeling process. The primary aim is to reduce lag time without significantly effecting the model's accuracy.

Finally, interactive knowledge acquisition is an area that would benefit from incremental learning techniques. For example, Bareiss and Porter (1987) have developed an expert `apprentice' in the area of clinical audiology. Their PROTOS system represents a novel approach to building expert systems, for it combines aspects of traditional knowledge acquisition techniques and automatic induction approaches (Michalski & Chilausky, 1981; Quinlan, 1986). Abstracting somewhat, PROTOS is an incremental learner of a DAG-structured knowledge base. Currently PROTOS relies heavily on the expert to define the organization of knowledge, though there are a few internalized principles of knowledge base design that allow it to operate in domains where expertise is harder to come by. Of pragmatic and psychological interest is the evolution of the basic level during an expert's training. Psychological studies indicate that there is a basic level of classification where performance on certain inference tasks is maximized. Basic level studies motivated many of the design decisions of COBWEB (e.g., category utility) and they may suggest important principles for classification-based expert systems. Our research on extending COBWEB to construct DAG knowledge bases will enable us to address this issue.


next up previous contents
Next: Frameworks for understanding incremental Up: Plan for Continued Research Previous: Domain-independent incremental methods

Douglas H. Fisher
Thu Jan 16 18:31:47 CST 1997