Facilitating Self-Education by Questioning Assumptive Reasoning

Robert G. FarrellAAAI 1988

Abstract

Traditional Intelligent Tutoring Systems (ITS) often focus on identifying "bugs" in a student's logic—errors where the student's output deviates from a pre-defined correct path. This paper introduces DECIDER, a Self-Education Aid (SEA) that shifts the focus from correctness to the underlying assumptions of the learner. By utilizing qualitative causal models and historical databases, DECIDER identifies the implicit logic in a student's foreign policy plan and presents "paradigmatic cases" where that same logic historically failed.

Introduction: Beyond "Bug-Hunting"

In the late 1980s, the dominant paradigm for educational software was the "debugger" model. If a student's answer didn't match the system's goal, it was flagged as an error. However, this approach often ignores the fact that students produce logical solutions based on specific, often implicit, assumptions. The core challenge is "blindness"—a state where a learner's mental model is so focused on a specific goal that they overlook complex interactions or potential violations of their assumptions. DECIDER seeks to bridge this gap by uncovering these assumptions and challenging them with historical reality.

Methodology: The DECIDER System

DECIDER functions as a Self-Education Aid (SEA) rather than a simple tutor. Its architecture is built around three core processes: model recovery, assumption questioning, and situated learning.

POLICY Models

At the heart of DECIDER are POLICY models, which are qualitative political models used to represent reasoning neutrally. These models consist of three primary components:

  1. Conditions: The qualitative states required for a plan (e.g., the existence of an insurgency).
  2. Inferences: Causal chains that explain how specific plans achieve these conditions.
  3. Prototypes: Typical objects (e.g., "Guns"), actions, and states that populate the model.
Formal components of a POLICY model, including conditions, inferences, and prototype definitions.

Model Recovery and Assumption Questioning

When a student proposes an action—such as "Have the US marines invade Managua"—DECIDER does not simply look for a keyword match. Instead, it uses its library of POLICY models to perform Model Recovery, building a causal graph that connects the specific action to high-level strategic goals.

Once the system understands the student's "idealized model," it searches for paradigmatic cases. These are historical events where the student's assumed causal link was broken. For instance, if a student assumes an invasion will succeed purely due to superior force, DECIDER might identify a case where resources were diverted to a different conflict, leading to failure.

A causal graph illustrating the hierarchical relationship between goals, subgoals, and potential failure modes like 'diversion'.

Situated Learning

To ensure these lessons are impactful, DECIDER utilizes a customized videodisc system. Instead of dry textbook descriptions, it presents "attached" views of history—such as dramatic footage from the Vietnam War—to make the consequences of failed assumptions more salient to the learner.

Data and Knowledge Engineering

As a classic AI project from 1988, DECIDER relies on curated knowledge rather than modern large-scale datasets. It utilizes:

  • Input Situation Database: Facts regarding contemporary crises (e.g., Nicaragua, Panama, Gaza).
  • Historical Case Database: A curated set of historical successes and failures indexed by the POLICY models they exemplify.
  • Expert Knowledge: The system's logic is informed by interviews with expert historians who identified the "paradigmatic examples" and common counter-arguments used against specific policies.

Key Insights

The most significant contribution of this work is the shift from "bug-fixing" to reasoning expansion. By treating student "errors" as logical deductions from limited assumptions, the system encourages a "coupled" learning environment. In this vision, the student and computer engage in a mutual argument, where the student is an active participant capable of disagreeing with the system's explanatory models.

Furthermore, the paper highlights the massive knowledge-engineering challenge inherent in this approach. Expert historians possess a familiarity with roughly one million cases; mimicking this level of expertise requires a significant investment in indexing and modeling historical knowledge.

Results and Limitations

DECIDER serves as a qualitative proof-of-concept. It successfully demonstrates the ability to map specific student plans (e.g., an invasion) to high-level policies through intermediate models. Crucially, it can link modern scenarios to distant historical parallels—such as connecting a plan for Nicaragua to the 16th-century failure of King Phillip II in the Netherlands—to illustrate how divided attention can undermine superior force.

However, the system faces a significant knowledge engineering bottleneck. Its effectiveness is strictly limited by the size and indexing of its case database. Additionally, the system initializes with certain prototypical goals (like maintaining influence) that may not align with every student's worldview, potentially introducing its own set of biases into the educational dialogue.

Summary updated

BibTeX
@inproceedings{farrell1988facilitating,
  author = {Robert G. Farrell},
  title = {{Facilitating Self-Education by Questioning Assumptive Reasoning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure