Abstract
Traditional Intelligent Tutoring Systems (ITS) often suffer from rigidity, intervening whenever a student deviates from a predefined expert path. This paper presents situated plan attribution, a flexible approach implemented in the REACT tutor. By focusing on "impasse points"—moments where environmental constraints or goal failures halt progress—the system provides targeted, context-aware guidance. Evaluation shows that this method allows students to resolve difficulties up to ten times faster than untutored peers while maintaining a non-intrusive learning environment.
Introduction
In the early 1990s, the "model tracing" paradigm dominated Intelligent Tutoring Systems. While effective in constrained domains, it struggled with real-world complexity where multiple valid paths exist or environmental states change dynamically. The authors identify three primary flaws in traditional plan recognition: rigidity in pathing, under-constrained modeling of mental states, and a lack of focus that leads to unnecessary interruptions.
To address these, the authors propose a "situated" approach. Instead of constant monitoring, the tutor intervenes only when the student reaches an impasse, using the current state of the environment to disambiguate the student's intent and provide relevant feedback.
Methodology: The REACT Architecture
The proposed system, REACT, is designed to train deep space communications operators. It is built upon the Soar cognitive architecture, which utilizes a hierarchical problem space to represent tasks and goals.
Impasse-Driven Tutoring
The core innovation of REACT is its focus on three specific types of "impasses":
- Action Constraint Impasses: Occur when a student attempts a valid step for which the environment is not yet prepared (e.g., trying to record before powering on the device).
- Goal Failure Impasses: Occur when a plan is completed but the desired environmental state was not achieved.
- Plan Dependency Impasses: Occur when a student initiates a new task before completing a mandatory predecessor.
When an impasse is detected, REACT employs an internal "expert" model to simulate the current situation, identify the root cause, and generate a natural language explanation alongside a repair strategy.
Evaluation and Results
The system was evaluated through a pilot study involving seven students using the Link Monitor and Control (LMC) simulator. Students were divided into a tutored group (using REACT) and a control group.
The results demonstrated significant advantages for the situated approach:
- High Precision: REACT correctly identified 58 distinct impasses across 604 actions with zero misinterpretations.
- Learning Efficiency: Tutored students resolved impasses and acquired new operational knowledge approximately ten times faster than the control group.
- Error Prevention: REACT effectively broke "failure cycles" where students would repeatedly miss goals without realizing it. By flagging "Goal Failures" immediately, the system ensured students corrected their mental models in real-time.
Conclusion
Situated plan attribution shifts the focus of tutoring from constant surveillance to strategic intervention at "teachable moments." By leveraging environmental context and the Soar architecture's problem-solving capabilities, REACT provides a model for more flexible, less intrusive, and highly effective intelligent instruction. Future work involves scaling this logic to even more complex domains where plan ambiguity is higher.
Summary updated