Intelligent Case Selection and Presentation

Robert G. FarrellIJCAI 1987

Abstract

Traditional educational AI systems often focus on the delivery of static content, failing to account for the dynamic nature of a learner's evolving understanding. This paper presents DECIDER, an intelligent tutoring system designed to teach complex decision-making in domains where answers are rarely binary, such as national security or high-level management. By utilizing "goal trees" to model a student's priorities and strategies, DECIDER algorithmically selects and presents historical cases that specifically target gaps in the student's mental model. This approach shifts the focus from what to teach to how and when to present information, facilitating a more effective "learning through experience" curriculum.

Introduction

In complex, multi-faceted domains, students must learn to navigate competing goals and prioritize outcomes rather than simply memorizing facts. A significant challenge in AI-driven education is the lack of pedagogical intuition—the ability to choose the right information to present at the right time. Most traditional systems are static, presenting the same cases regardless of the learner's background.

Robert Farrell of Yale University addresses this by introducing DECIDER. The system is built on the premise that a "good teacher" adapts their presentation based on a model of the audience. DECIDER automates this intuition by using a dynamic student model to drive case selection from a large library, ensuring that each historical scenario serves a specific pedagogical purpose.

Methodology and System Architecture

The DECIDER architecture relies on the interplay between three core components that represent knowledge and the teaching process:

  1. The Student Goal Tree (SGT): Based on Carbonell’s goal trees, this represents the student’s current understanding of goals, subgoals (S-links), and the relative importance (RI-links) between them.
  2. The Possible Goal Tree (PGT): This serves as the "expert" or "teacher" model—the ideal framework the student is intended to approximate.
  3. The Teacher-Initiative Algorithm: A rule-based engine that selects the next case based on a specific pedagogical hierarchy:
  4. Introduce Relative Importance: Presenting cases where a student must choose between two known goals.
  5. Introduce Subgoals: Demonstrating how a known goal serves a higher-level objective.
  6. Introduce New Goals: Incorporating entirely new concepts into the student's model.
  7. Refine Specificity: Moving from abstract concepts to nuanced, specific goals.

Cases in the DECIDER database are indexed by "Pedagogical Opportunities." Rather than simple keyword tagging, cases are marked by their ability to demonstrate goal failures, plan successes, or the necessary sacrifice of one priority for another.

Learning Through Sacrifice and Functional Indexing

A central insight of the DECIDER system is the concept of learning via sacrifice. The system assumes that students truly grasp the "relative importance" of goals only when forced to select a plan that achieves one objective at the expense of another. This forces the student to reconcile their internal priorities with historical realities.

Furthermore, the paper posits that the primary difference between a novice and an expert lies in memory organization. While novices often index information by emotional import or surface-level features, experts index by functional categories such as plans, failures, and goals. DECIDER encourages this transition by presenting cases through a functional lens, effectively training the student to adopt an expert's indexing strategy.

Implementation and Results

The system was implemented using a database of U.S. national-security decision-making cases, including scenarios from the Reagan administration where military objectives were constrained by civilian casualty concerns. While the knowledge base used for testing was specific to national security, the architecture is domain-independent; the tutorial engine is decoupled from the database, allowing it to be applied to any decision-making field.

As a system architecture paper, the results are primarily qualitative. The author demonstrates that DECIDER successfully:

  • Provides Dynamic Scaffolding: The SGT expands over multiple rounds, moving from simple goal recognition to complex priority management.
  • Automates Curation: The algorithm effectively filters large libraries to find the specific scenario that addresses the student's current "frontier" of knowledge.
  • Facilitates Failure-Driven Learning: By allowing students to select plans and then presenting the historical outcome, the system creates an environment where students must explain and learn from the discrepancy between their expectations and reality.

Limitations and Future Directions

Despite its innovative approach, DECIDER operates on a simplified, discretized model of decision-making. It treats decisions as isolated events made by a single actor, whereas real-world decisions are often continuous, multi-agent processes involving negotiation. Additionally, the paper lacks a formal user study to quantitatively compare DECIDER against traditional teaching methods. Future work involves developing more sophisticated monitoring of student interactions to achieve an "invisible" pedagogy, where the system adapts so subtly that the student is unaware of the underlying curriculum steering.

Summary updated

BibTeX
@inproceedings{farrell1987intelligent,
  author = {Robert G. Farrell},
  title = {{Intelligent Case Selection and Presentation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1987},
}

Loading related content...

Figure