Reasoning About Assumptions in Graphs of Models

Sanjaya Addanki, Roberto Cremonini, J. Scott PenberthyIJCAI 1989

Abstract

This paper introduces the Graphs of Models (GoM) paradigm, a framework for organizing vast amounts of physical domain knowledge into a network of simplified, assumption-based models. When a chosen model fails to predict real-world behavior, the system uses qualitative reasoning to automatically transition to a more complex, accurate model within the graph. This approach mimics human engineering strategies, prioritizing simplicity and computational efficiency over universal precision.

Introduction

In the late 1980s, AI researchers faced a significant hurdle known as the "Knowledge Representation Problem." Attempting to represent every law of physics and material property in a single, monolithic knowledge base made reasoning computationally expensive, often leading to exponential complexity. However, human experts rarely reason from first principles for every task; instead, they utilize simplified models by making specific assumptions—such as ignoring air resistance or assuming a rigid frame. The challenge addressed here is the automation of this strategy: decomposing a domain into manageable models, formally representing their underlying assumptions, and intelligently switching models when predictions fail.

The Graphs of Models Paradigm

The authors propose the Graphs of Models paradigm, where nodes represent individual models and directed edges represent transitions between assumptions. Each model consists of two primary components: a set of rules for domain analysis (e.g., $F=ma$) and a set of rules for changing assumptions.

A Graph of Models showing transitions between different physical assumptions.

Models are categorized as "materialized" (pre-defined in the graph) or "non-materialized" (potential combinations of assumptions). The system aims to find the simplest materialized model that provides an acceptable approximation of the observed world. By treating assumption changes as actions with specific post-conditions, model selection is transformed into a search problem solvable via standard AI inference techniques.

Qualitative Reasoning and Influence Nets

When a model's prediction deviates from measured values—a state termed an "Empirical Conflict"—the system must determine how to adjust its assumptions. Rather than performing high-precision recalculations, GoM utilizes a qualitative "(+, -, 0) calculus" known as delta-vectors. If a predicted acceleration is higher than the measured value, the system generates a delta-vector of (acceleration, -), indicating a need for a model where acceleration decreases.

To facilitate this, the framework employs Influence Nets (I-Nets). These are semantic networks that represent how various parameters affect one another qualitatively.

An Influence Net (I-Net) representing qualitative relationships between physical parameters.

As shown in the I-Net, parameters like torque ($Tau$) and moment of inertia ($I$) have defined relationships with angular acceleration ($\alpha$). These networks allow the system to back-propagate required changes through complex equations to identify which assumptions (e.g., switching from "no friction" to "Coulomb friction") will yield the desired parameter shift.

Experimental Validation

The paradigm was validated through custom implementations in four physical domains: mechanics (gear train transmissions), thermodynamics (gas behaviors), fluids (laminar vs. turbulent flow), and geometric structure.

A primary benchmark involved a gear train transmission. Initially, the system employed a simplistic frictionless model. When this model failed to match experimental data, the system detected the conflict, utilized I-Nets to determine that adding friction would reconcile the discrepancy, and automatically transitioned to a more complex model incorporating Coulomb friction. This demonstrated the system's ability to avoid the overhead of a "universal" physics engine by reasoning over localized, model-specific knowledge bases.

Limitations and Future Work

While GoM provides a robust framework for model switching, several limitations remain. The system currently understands the direction of change but not the magnitude; it cannot yet distinguish between a negligible amount of friction and a significant one. Furthermore, the current "goodness" measure for model selection is a simplistic numerical sum of satisfied delta-vectors. Finally, scaling the system to handle the NP-complete problem of searching all possible non-materialized assumption combinations remains a significant challenge for future research.

Summary updated

BibTeX
@inproceedings{addanki1989reasoning,
  author = {Sanjaya Addanki and Roberto Cremonini and J. Scott Penberthy},
  title = {{Reasoning About Assumptions in Graphs of Models}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1989},
}
Figure