Behavioral Aggregation Within Complex Situations: A Case Study Involving Dynamic Equilibria

Shankar A. Rajamoney, Sang Hoe KooAAAI 1991

Introduction

Presented at AAAI-91, this research addresses a fundamental bottleneck in artificial intelligence: the computational complexity of reasoning about complex physical systems. While Qualitative Reasoning (QR) provides a powerful framework for predicting system behavior without precise numerical data, it often suffers from a "state space explosion" when applied to large-scale models. The authors propose a "behavioral aggregation" framework that simplifies these models by grouping entities in dynamic equilibrium into single, higher-level units.

The Problem: State Space Explosion

Qualitative Reasoning aims to describe physical phenomena using symbolic values (e.g., "increasing," "decreasing," "zero"). However, as the number of components in a system grows, the number of possible qualitative states increases exponentially. For instance, modeling every valve, pipe, and reaction in a chemical plant from first principles leads to a proliferation of ambiguous states that are mathematically intractable. The authors observe that human experts avoid this by using abstraction—treating complex sub-systems as single functional blocks rather than a collection of parts.

The Behavioral Aggregation Framework

The core of the proposed method is Behavioral Aggregation, which treats groups of entities that maintain a dynamic equilibrium as a single unit. This framework is built on three pillars:

  1. Composability: Aggregates are modular and can be nested into larger structures.
  2. Uniformity: Aggregates are represented using the same Qualitative Process (QP) Theory language as individual components, ensuring compatibility with existing simulators.
  3. Opacity: The internal "micro-steps" of an aggregate are hidden, allowing the reasoner to focus on high-level behavior without tracking every internal interaction.
Figure 1: Conceptual model of behavioral aggregation showing internal compensation in response to external gas flow between two containers.

The technical implementation relies on modeling External Stresses and Internal Compensatory Processes. When an external influence (such as a gas leak) disturbs the equilibrium, the aggregate model triggers an internal process (like increased evaporation) to restore balance. This allows the simulator to skip the micro-level fluctuations and jump directly to the high-level result.

Figure 3: Formal representation of an aggregate in Qualitative Process Theory, including individual views, negative-stress conditions, and compensation processes.

Experimental Evaluation

The framework was evaluated using the Qualitative Process Engine (QPE) across several fluid-domain scenarios. The results demonstrated significant reductions in complexity:

  • Scenario C (Two-container system): The aggregate model reduced the state space from 70 states to just 13, resulting in a nearly 4x speedup (from 111.9 to 32.5 seconds).
  • Scenario D (Complex phase-change system): A detailed first-principles simulation failed to complete within 1,000 CPU seconds. In contrast, the aggregate version successfully generated a compact graph of only 16 states and 60 transitions.

Key Insights

  • Behavioral vs. Structural Glue: Aggregation is driven by how entities act together (e.g., maintaining equilibrium) rather than just physical proximity. The "glue" is the collective interaction that results in a predictable response to the environment.
  • Equilibrium as a Simplifier: By assuming a system remains in equilibrium, the reasoner can suppress internal process rates and focus on net changes, drastically reducing the number of variables the computer must track.
  • Hierarchical Scaling: The uniformity of the representation allows for "aggregates of aggregates," mirroring the hierarchical abstraction used by human engineers to manage complexity.

Conclusion and Future Work

Behavioral aggregation transforms intractable qualitative models into interpretable, compact descriptions that align with human expertise. However, the current approach requires manual identification of aggregates by a user. Future research directions include automating the discovery of these aggregates and extending the framework to handle other complex behaviors like oscillations, cycles, and feedback loops.

Summary updated

BibTeX
@inproceedings{rajamoney1991behavioral,
  author = {Shankar A. Rajamoney and Sang Hoe Koo},
  title = {{Behavioral Aggregation Within Complex Situations: A Case Study Involving Dynamic Equilibria}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure