Understanding Behavior Using Consolidation

Tom Bylander, B. ChandrasekaranIJCAI 1985

Abstract

This paper introduces consolidation, a novel framework for reasoning about physical device behavior by composing individual component actions into higher-level "composites." By moving away from global simulations toward localized causal patterns, consolidation offers a computationally efficient model that aligns with human intuitive understanding of machines.

Introduction

In the mid-1980s, the field of "Naive Physics" aimed to provide AI with commonsense reasoning about the physical world. The dominant paradigm, Qualitative Simulation (QS), faced significant hurdles: extreme computational complexity due to its global nature, a lack of clear causal explanations, and difficulty in supporting multi-level hierarchical reasoning.

To address these issues, we propose Consolidation. Instead of simulating an entire system simultaneously, consolidation infers the behavior of component groups (substructures) and treats them as new, singular "composite" components. This localized approach allows for scalable reasoning and clearer causal tracing.

Methodology: Structural and Behavioral Primitives

Consolidation relies on a set of primitives to describe physical systems. Structural primitives define physical relationships such as Connection (contact) and Containment (holding substances). Behavioral primitives consist of a small vocabulary of "verbs" describing component actions: Allow (passage), Influence (pumping/expelling), Move, Create, and Destroy.

Behavioral primitives applied to a light bulb circuit.

The core of the method is the application of Causal Patterns—logical rules of composition. For instance, a "Pump" behavior in series with an "Allow" behavior can be consolidated into a larger "Pump" behavior across the entire path. This hierarchical inference hides internal complexity, exposing only the outward behavior of the composite to the rest of the system.

Case Studies and Validation

The theory is validated through two classic models: a light bulb circuit and a water container system.

The Light Bulb Circuit

In the circuit model (see Figure 2), the system derives that "light is created" when the switch is closed. It achieves this by composing the battery’s electrical "pump" behavior with the bulb’s "resistance" and "light-creation" behaviors. The explanation for the light being on is the direct trace of this consolidation process.

Water and Heat Transfer

The second model involves two tanks at different temperatures connected by a pipe.

Behavioral primitives for a water container system.

The model demonstrates how a "move water" behavior triggers a "move heat" behavior, eventually affecting the "expel heat" (temperature) behavior of the destination tank. This illustrates the framework's ability to handle the movement and transformation of "substances" like water and heat.

Key Insights

  • Locality: Reasoning occurs in well-defined, locally contained steps, avoiding the need for global knowledge of the entire system.
  • Substance-Centric Behavior: Device interaction is primarily viewed as the movement and transformation of substances (electricity, water, heat, light).
  • Causality as Composition: The "why" behind a device's function is provided by the trace of the consolidation process itself.

Conclusion and Future Work

Consolidation provides a path toward reasoning agents that avoid the "state-space explosion" typical of 1980s AI. However, challenges remain, including the combinatorial explosion of states in complex composites and the need for a more robust library of patterns to handle momentum, temporality, and multi-substance interactions. Future work will focus on refining these causal patterns and completing the implementation of the reasoning engine.

Summary updated

BibTeX
@inproceedings{bylander1985understanding,
  author = {Tom Bylander and B. Chandrasekaran},
  title = {{Understanding Behavior Using Consolidation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure