FITS: A Fraction Intelligent Tutoring System

Hyacinth S. NwanaAAAI 1991

Introduction

In the early 1990s, the field of Intelligent Tutoring Systems (ITS) primarily focused on "primitive" domains, such as simple integer arithmetic. These systems often relied on linear step-by-step models that failed to capture the complexity of real-world learning. FITS (Fraction Intelligent Tutoring System) was developed to address the domain of fraction addition—a task requiring a hierarchy of sub-skills, including integer multiplication, finding the lowest common denominator (LCD), and simplification.

The challenge in tutoring fractions lies not just in the procedural steps, but in the strategic application of knowledge. Students frequently possess the necessary sub-skills but struggle with the "when" and "where" of their application. FITS addresses these complexities through a flexible, modular architecture that supports idiosyncratic solution paths.

System Architecture

FITS was implemented using Quintus Prolog on SUN and VAX workstations. Its design emphasizes a strict separation between domain-specific knowledge and domain-independent tutoring logic. This modularity allows for greater portability and easier refinement of the system's pedagogical strategies.

The modular architecture of the FITS system, illustrating the flow of control and data between the user interface, student models, and domain-specific knowledge bases.

As shown in the architecture diagram, the system is composed of several key modules:

  • User Interface Module: Manages the interaction with the student, incorporating a Problem Solving Monitor and a Student Modelling Module.
  • The Bug Catalogue: A curated library of common student errors (malrules) derived from empirical research. This is used for diagnostic purposes to identify specific misconceptions.
  • Student-Tutor History: A dynamic record of all interactions, enabling the system to distinguish between accidental "slips" and fundamental conceptual misunderstandings.
  • Tutoring Strategy Module: A rule-based engine that determines the system's next action—such as providing a hint or showing an example—by comparing the current Student Model with an Ideal Student Model.
  • Domain Expert Rulebase: Contains the procedural knowledge required to solve fraction problems correctly.

Instructional Design and Insights

A core innovation of FITS is the explicit separation of skill and strategy. The system recognizes that a student might be proficient in finding an LCD but may fail a problem because they do not realize it is required at a specific juncture. To address this, FITS utilizes "Vanlehn-style" teaching sessions to communicate the strategic "plan" of fraction operations.

Furthermore, FITS avoids the rigidity of contemporary "model-tracing" tutors by supporting idiosyncratic paths. For instance, students are permitted to find equivalent fractions before or after processing whole numbers in mixed-fraction problems. The system also employs a "Good Student Privilege" mechanism: proficient students can skip intermediate steps, but if errors occur, the system dynamically reverts to a step-by-step mode to facilitate diagnosis.

Evaluation and Results

The system's development relied on pre-test responses to initialize student models and the Bug Catalogue for error diagnosis. While quantitative evaluation data was published separately, qualitative analysis of student protocols (e.g., interactions with a student named "Rose") demonstrated the system's functional sophistication. FITS successfully identified classic fraction errors, such as adding denominators, and provided context-aware feedback.

Beyond simple correctness, FITS provides efficiency feedback. If a student completes a problem using a "long-winded" method, the system displays a more optimal solution path to encourage procedural efficiency. The authors claim that this architecture successfully integrates and improves upon the design principles of major contemporary models, including Anderson’s Geometry tutor.

Limitations and Future Directions

Despite its modular strengths, FITS reflects the technological constraints of the early 1990s. The system lacks natural language processing, resulting in formulaic outputs. The interaction remains largely "system-led," with the tutor dominating the dialogue flow. Additionally, the tutoring strategies are static and do not evolve through machine learning. Future work in this area involves improving the "naturalness" of the dialogue and tailoring explanations to more granular student profiles.

Summary updated

BibTeX
@inproceedings{nwana1991fits,
  author = {Hyacinth S. Nwana},
  title = {{FITS: A Fraction Intelligent Tutoring System}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure