Plan Recognition Strategies in Student Modeling: Prediction and Description

This paper describes the student modeler of the GUIDON2 tutor, which understands plans by a dual search strategy. It first produces multiple predictions of student behavior by a model-driven simulation of the expert. Focused, data-driven searches then explain incongruities. By supplementing each other, these methods lead to an efficient and robust plan understander for a complex domain. 1. Basic problem: Modeling strategic problem solving Diagnostic problem-solving requires domain knowledge and a plan for applying that knowledge to the problem. A hypothesis-directed diagnostic plan is a rationale for focusing on diagnoses (partial solutions) and for gathering data to solve the problem. The plan is thus a strategy for selecting and ordering the application of domain knowledge. Teaching diagnosis involves recognizing the intent behind a student’s behavior, so that missing knowledge can be distinguished from inappropriate strategies. The teacher interprets behavior, critiques it, and provides advice about other approaches. To do this successfully and efficiently in a complex . domain, the teacher benefits from multiple, complementary modeling strategies. GUIDON2 is a tutoring program that uses the case method approach to teach medical diagnosis [5]. The system divides this task among three components: an “expert,” a student modeler, and an instructional manager (see Figure l-l). Its expert component, NEOMYCIN [4], separately and explicitly represents knowlcdgc about the medical domain and the domain-independent strategies of diagnostis. The student modeler, a subprogram called IMAGE, interprets the student’s behavior by using NEOMYCIN’s knowledge, evalu;ltcs the student’s skill, and produces alternatives. The instructional module of GUIDON2 will then apply discourse and teaching strategies in deciding whether to interrogate or advise the student. . A model of student strategies in medical diagnosis must disambiguate the possible purposes and knowlcdgc underlying the student’s actions. ‘1’1~~ approaches followed by other plan recognizcrs and student modclcrs arc not suffcicnt hcrc because: (1) the complex domain makes thorough searches impractical, whether top-down or bottom-up; (2) we are not modeling only facts and rules used in isolation, but also the procedures for applying them; (3) every one of the student’s actions must be monitored in case the teaching module decides to interrupt; (4) his behavior must be cvaluatcd and not just explained; and (5) WC might not have any explicit goal statcmcnts from the student, so WC expect to rely only on his qucrics for problem data as cvidcncc for his thinking. Y i