Setting up Large-Scale Qualitative Models

Brian Falkenhainer, Kenneth D. ForbusAAAI 1988

Abstract

Qualitative Physics initially struggled with scaling from simple toy models to complex industrial systems. This paper by Falkenhainer and Forbus (1988) addresses the "scaling crisis" by introducing a formal framework for modeling assumptions. By explicitly defining simplifying and operating assumptions, the authors enable the automatic extraction of minimal models tailored to specific queries. This approach is demonstrated on a large-scale model of a Navy steam propulsion plant, significantly reducing computational complexity and enabling efficient qualitative reasoning.

Introduction

In the late 1980s, the field of Qualitative Physics reached a bottleneck. While researchers could successfully model simple systems like coffee pots or basic circuits, industrial-scale systems—such as a Navy propulsion plant—remained computationally intractable. The state-space explosion inherent in modeling every possible interaction in a complex system overwhelmed existing tools. This seminal paper introduced a formal methodology for managing model complexity by telling a computer what to ignore, marking a significant advancement in qualitative reasoning.

The Challenge of Scale

Moving from toy problems to real-world engineering systems presents three primary hurdles:

  1. Organization: Structuring models that span multiple levels of abstraction, from high-level functional blocks to microscopic details.
  2. Relevance: Avoiding the simulation of every possible state, which leads to exponential growth in the state space.
  3. Selection: Identifying the specific subset of physics necessary to answer a particular question (e.g., "Is the boiler efficient?") without involving irrelevant components.

The authors argue that effective modeling requires meta-knowledge: the system must understand the assumptions under which specific physical laws are applicable.

Formalizing Modeling Assumptions

The core contribution is the introduction of explicit assumptions to control model scope and behavior:

  • Simplifying Assumptions: These define the "grain size" and "perspective" of the model. For instance, a CONSIDER statement can toggle between viewing a system as a single functional unit or a collection of parts, or focus exclusively on thermal versus volumetric properties.
  • Operating Assumptions: These describe default modes, such as "steady-state" or "normal operation," allowing the system to ignore transient behaviors that are irrelevant to the current query.
Hierarchical perspectives and modeling assumptions in a qualitative model.

To operationalize these, the authors developed a selection algorithm using an Assumption-based Truth Maintenance System (ATMS). The algorithm identifies the minimal set of simplifying assumptions required to justify a specific query, effectively pruning the model to its most relevant components.

Case Study: The Navy Steam Propulsion Plant

The framework was validated using a domain model of a Navy steam propulsion plant, which was the largest qualitative model of its time. The model included 8 object types, 23 views, and 14 processes, resulting in 79 quantities and over 8,600 logical clauses upon instantiation.

Schematic of the Navy steam propulsion plant used as a case study.

Results and Efficiency

The use of assumption-based selection transformed the modeling process. Without these techniques, a "total envisionment" of the plant would crash contemporary hardware. By focusing on specific queries, the system reduced simulation times to just a few minutes. It successfully answered complex questions regarding plant efficiency and fault diagnosis (e.g., causes of black smoke in the furnace) by instantiating only 15–25 relevant quantities out of thousands.

Conclusion and Future Work

While the paper successfully addresses model selection, it highlights remaining challenges. Automating the "structural abstraction" from raw CAD data to functional models remains difficult. Additionally, while simplifying assumptions are well-handled, the automated selection of operating assumptions (e.g., determining when steady-state applies) requires further research. Finally, the lack of deep geometric reasoning limits the model's fidelity for certain engineering tasks.

Summary updated

BibTeX
@inproceedings{falkenhainer1988setting,
  author = {Brian Falkenhainer and Kenneth D. Forbus},
  title = {{Setting up Large-Scale Qualitative Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}

Loading related content...

Figure