Abstract
This paper presents a plan-based approach to student modeling within Intelligent Computer Assisted Instruction (ICAI) systems. By conceptualizing mathematical problem-solving as a hierarchical planning process, the proposed system can infer hidden reasoning steps, accommodate multiple valid solution paths, and systematically categorize procedural errors in arithmetic.
Introduction
Traditional educational software often relies on simple "answer-key" matching, which fails to capture the complexity of human learning. This research addresses three primary challenges in teaching elementary algebra and arithmetic:
- Non-determinism: Multiple correct paths often exist for a single problem (e.g., simplifying a fraction before or after addition).
- Elliptic Solutions: Students frequently omit intermediate steps, making their internal reasoning opaque to the system.
- Stereotyped Errors: Mistakes are often not random but result from the consistent application of "buggy" rules.
The objective is to move beyond final-answer grading toward a "global model" of student competence that informs personalized pedagogical feedback.
System Architecture and Methods
The proposed architecture centers on a Plan Inference engine. The system is structured around several key modules: the Expert, the Modelizer, the Pedagogue, and the Student Model.
The core data structure is the Context Frame, representing specific mathematical tasks. Key technical components include:
- Hierarchical Task Analysis: Calculations are decomposed into obligatory and optional tasks.
- Multi-Level Method Modeling: Methods are tracked across four levels of "memory," ranging from frequently used (Short-term) to never used.
- Search Heuristics: Three tiered heuristics (Regularity, Extended, and Universal) are used to find the most plausible plan behind a student's response.
- Dialogue-Driven Disambiguation: When multiple plans explain an answer, the system initiates a targeted dialogue to clarify the student's logic.
Key Insights: Calculations as Plans
A central innovation is treating mathematical calculations as goal-oriented plans rather than mere sequences of numbers. This allows the system to use AI planning techniques to "fill in the blanks" of a student's work.
Ambiguity is viewed as a pedagogical opportunity. As shown in Figure 4, if a student provides a result like 29/21 for 4/6 + 5/7, the system can infer two distinct valid subplans: one involving early simplification and one involving late simplification. This ambiguity triggers a dialogue, forcing the student to externalize their hidden steps and allowing the system to update the Student Model dynamically.
Key Results and Discussion
The functional prototype successfully identifies specific "buggy" procedures (e.g., adding numerators and denominators directly) and distinguishes between different valid solution orders. By formalizing errors as "illegitimate tasks," the system can differentiate between simple slips and fundamental conceptual misunderstandings.
However, limitations remain. The system currently focuses on single-step equalities, and scaling plan inference to more complex domains like calculus or physics presents significant computational challenges. Furthermore, the authors note that these "plans" are descriptive models of actions and may not perfectly reflect a student's internal mental state.
Conclusion
By integrating plan inference into student modeling, this research provides a more nuanced framework for ICAI. It moves the field toward systems that can truly "understand" a student's reasoning process, paving the way for more effective and interactive digital tutors.
Summary updated