A Representation for Complex Physical Domains

Sanjaya Addanki, Ernest DavisIJCAI 1985

Introduction

In the mid-1980s, AI researchers faced a significant challenge in physical reasoning. Systems were typically divided into two camps: "naive physics" models that understood the world qualitatively (e.g., tipping a cup causes a spill) and "expert systems" that relied on complex differential equations. While qualitative models were too shallow for engineering tasks, quantitative models were often too computationally expensive for general problem-solving.

Sanjaya Addanki and Ernest Davis introduced PROMPT (Physical Reasoning by the Organisation and Modification of ProtoTypes) to bridge this gap. PROMPT provides a representation that supports both high-level heuristic reasoning and low-level physical analysis, enabling AI to answer complex engineering questions: How do I design a pen that writes upside down? Why won't my pen write on glass? What happens to a pen if the temperature hits 200°F?

The PROMPT System

The core of the PROMPT system is built upon two pillars: a sophisticated ontology and a hierarchical library of prototypes.

1. A Sophisticated Ontology

The authors expanded the AI's view of time and change. Rather than a simple sequence, time is defined through "chronicles" (complete world histories) and "histories" (functions describing specific events). This allows the system to reason about simultaneous processes, such as a resistor simultaneously conducting electricity and generating heat.

2. Prototypes

Prototypes are structured chunks of information about entities, relations, or processes. Organized into IS-A (specialization) and PART-OF (composition) hierarchies, they serve as starting points for reasoning. Instead of searching through every law of physics, the system uses a prototype—like a ball-point pen—and modifies it to solve specific problems, such as designing a pen for zero-gravity environments.

Case Study: The Pen World

To demonstrate the system's capabilities, the authors introduced the "Pen World," a rigorous physical breakdown of a standard ball-point pen. This benchmark models the interactions of gravity, viscosity, cohesion, adhesion, and friction across seven distinct parts, including the housing, cartridge, and tip.

Figure 1: Details of the ball-point pen tip, illustrating the components modeled in the Pen World benchmark.

By explicitly defining "continuation conditions" for processes, PROMPT can perform limit analysis. For example, it can predict that a pen will stop writing if the friction between the ball and the paper surface falls below the internal friction within the tip assembly.

Key Insights and Reasoning

PROMPT's effectiveness stems from several key conceptual shifts:

  • The Common Substrate: By placing first-principles reasoning and high-level schemas in the same hierarchy, the system can "zoom in" to deep physics only when high-level heuristics fail.
  • Essential Change: Processes are defined by their primary physical change (e.g., liquid changing position during flow). This allows continuous phenomena to be treated as discrete masses, simplifying the application of laws like momentum.
  • Diagnostic Reasoning: The ontology supports reasoning about alternative sequences of past states, which is essential for identifying why a device has failed.

Implementation and Results

Initial implementation of an "envisioning system" showed that the prototype structure significantly simplified the control needed for forward-chaining reasoning. Unlike previous systems confined to a "closed world" of pre-set rules, PROMPT could generate new transforms and modifications from first principles. This allowed it to reason at an "expert level" by exchanging information between parallel, weakly coupled analyzers to constrain the search space.

Limitations and Future Work

Despite its strengths, the authors identified several hurdles for future development:

  • Shape Representation: Representing complex 3D geometry remains difficult. While the system can model a pen as "cylindrical," it lacks a robust engine to determine if complex parts will physically fit together.
  • Numerical Integration: The current focus is on symbolic representation; integrating the numerical execution of complex fluid dynamics (like Navier-Stokes equations) is a target for future work.
  • Scaling: Moving beyond the controlled "Pen World" to complex systems like power plants would require a significantly larger library of prototypes.

Summary updated

BibTeX
@inproceedings{addanki1985representation,
  author = {Sanjaya Addanki and Ernest Davis},
  title = {{A Representation for Complex Physical Domains}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure