Abstract
This paper presents a formal framework for student modeling in Intelligent Tutoring Systems (ITS) that addresses the dynamic and nonmonotonic nature of human learning. By integrating a four-valued logic language (SMDL) with an Assumption-based Truth Maintenance System (ATMS), the authors develop the Hypothetical Student Model Inference System (HSMIS). This system is capable of representing a student's belief states—including ignorance and contradictions—and efficiently updating the model as the student's knowledge evolves.
Introduction
In the development of Intelligent Tutoring Systems, the "Student Model" serves as the system's internal representation of a learner's knowledge state. Traditional models, such as overlay or buggy models, often suffer from rigidity. They struggle to handle two fundamental aspects of human cognition: nonmonotonicity and the representation of the unknown. Learning is rarely a linear accumulation of facts; students often abandon old misconceptions in favor of new insights, requiring the system to retract previous assumptions. Furthermore, a student's inability to answer a question is a distinct state of knowledge that binary logic (True/False) cannot adequately capture.
Formalizing Knowledge with SMDL
To address these challenges, the authors introduced the Student Model Description Language (SMDL). SMDL extends standard Prolog by utilizing a four-valued logic system consisting of true, false, unknown, and fail. This allows the system to distinguish between a student believing a proposition is false and a student simply being unaware of it.
As shown in the truth table above, the inclusion of "unknown" (unk.) and "fail" allows for more nuanced logical operations. This formalization enables the system to deduce why a student cannot reach a specific conclusion, providing deeper diagnostic insights than traditional binary systems.
The HSMIS Architecture
The core engine of this framework is the Hypothetical Student Model Inference System (HSMIS). HSMIS treats student modeling as an inductive inference problem, essentially "debugging" its theory of the student's mind based on observed answers.
As illustrated in the architecture diagram, HSMIS incorporates several key components:
- ATMS Integration: By utilizing de Kleer’s Assumption-based Truth Maintenance System, HSMIS can maintain multiple hypothetical versions of a student's mental state. When a student provides a contradictory answer, the ATMS identifies the conflicting assumptions and facilitates a transition to a new, consistent model state.
- Virtual Oracles: To minimize the "cost" of modeling (i.e., over-interrogating the student), the system employs "virtual oracles." These generate plausible student responses based on an ideal teacher model until the student's actual behavior contradicts these assumptions.
Implementation and Results
The system was evaluated through implementation in two instructional domains: Geography (modeling plant growth conditions) and Chemical Reactions. Implemented in Common ESP (Extended Self-contained Prolog) on SPARC stations, the system demonstrated real-time performance in these domains.
Key results indicated that the system could successfully identify incorrect clauses in its student model and replace them when presented with new evidence (refutation logic). Because SMDL is based on general logic programming, the HSMIS engine remains domain-independent, capable of being applied to any subject matter representable via Prolog-like rules.
Limitations and Future Work
While robust, the system faces computational challenges when resolving massive contradictions, currently relying on domain-independent heuristics that may become expensive in complex domains. Furthermore, the current model assumes a student should be consistent. Future work aims to use the ATMS to model "multiple worlds" within a single student's mind, acknowledging that learners often hold contradictory beliefs simultaneously. The authors also suggest refining belief revision mechanisms to ensure more efficient, minimal updates to the student model.
Summary updated