The Decomposition of a Large Domain: Reasoning About Machines

Craig StanfillAAAI 1983

Introduction

In the early 1980s, artificial intelligence researchers faced a significant hurdle: "large domains." These were problems of such complexity that a flat, rule-based knowledge base became unwieldy and unmanageable. Reasoning about mechanical systems is a quintessential example, requiring a synthesis of geometry, kinematics, pneumatics, and dynamics. This 1983 paper introduces Mack, a system designed to bridge the gap between raw geometric data and physical intuition by decomposing the reasoning task into two distinct hierarchies.

The Dual-Hierarchy Architecture

The core contribution of Mack is its dual-hierarchy structure, which manages complexity by organizing both domain knowledge and the reasoning process itself.

The Semantic Hierarchy

The Semantic Hierarchy organizes knowledge into seven disjoint "sub-worlds." Each world—ranging from Algebra and Linear Geometry to Mechanics and Pneumatics—operates with its own primitives and rules. This encapsulation allows high-level experts to query lower-level ones without needing to understand their internal logic. For instance, the Pneumatics expert can request geometric calculations from the Solid Geometry expert without knowing the underlying algebraic implementation.

The Semantic Hierarchy organizes domain knowledge into specialized sub-worlds.

The Epistemological Hierarchy

Complementing the semantic structure is the Epistemological Hierarchy, which defines the sequence of increasingly abstract models the system generates to understand a machine.

The Epistemological Hierarchy outlines the stages of reasoning from static geometry to dynamic processes.

The reasoning progression includes:

  • Static Model: Raw shapes and spatial coordinates.
  • Kinematic Model: Constraints on motion derived from part contact.
  • Pneumatic Model: Identification of gas-containing chambers.
  • Force and Acceleration Models: Calculating forces based on pressure and linking them to motion changes.
  • Process Model: The final qualitative output describing cause-and-effect relationships (e.g., "Increasing pressure causes the piston to move right").

From Geometry to Qualitative Process

Unlike systems that start with predefined functional relationships, Mack treats geometry as the "source of truth." It begins with symbolic descriptions of parts—such as cylinders defined by points and boolean operations—and derives behavior through analysis.

A key insight of the system is its shift from precise numerical simulation to Qualitative Process Theory. Instead of solving complex differential equations, Mack captures the tendencies of variables. It identifies how variables like pressure, volume, and temperature influence one another, providing a human-like explanation of machine behavior.

Evaluation and Results

Mack was evaluated on several mechanical assemblies, most notably a piston and cylinder block. The system successfully navigated the entire pipeline, moving from a list of coordinates to a qualitative process model. Notably, Mack demonstrated automated parameterization, creating its own variables (such as X-1 for displacement) to describe the degrees of freedom it discovered.

In terms of performance, the piston example required approximately five minutes of computation on a VAX 11/780 with 4MB of memory. While modest by modern standards, this was a significant demonstration of multi-domain integration in 1983. Beyond pistons, the system also successfully "understood" bearings, tanks, and multi-bearing systems.

Discussion and Future Work

While Mack provided a robust framework for mechanical reasoning, the author noted several areas for expansion. The hierarchies could be deepened to include more granular mathematical structures or foundational logic. Furthermore, the system was limited to "simple" machines following ideal gas laws and Newtonian physics, leaving complex fluid dynamics and flexible materials for future research. Finally, the reliance on 750 hand-crafted rules highlighted the "knowledge acquisition bottleneck" that would eventually shift the field toward machine learning.

Summary updated

BibTeX
@inproceedings{stanfill1983decomposition,
  author = {Craig Stanfill},
  title = {{The Decomposition of a Large Domain: Reasoning About Machines}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}
Figure