Abstract
Qualitative Reasoning (QR) provides a framework for predicting physical system behavior using imprecise knowledge. However, early QR systems suffered from "state space explosion," limiting their applicability to complex engineering problems. This paper presents the Qualitative Physics Compiler (QPC), a hybrid architecture that decouples the logical task of model construction from the mathematical task of simulation. By utilizing a domain theory of model fragments and a specialized simulation algorithm (QSIM), QPC achieves massive performance improvements and successfully models large-scale systems, such as chemical processing plants with hundreds of variables.
Introduction
In many engineering and "common sense" scenarios, precise numerical data or complete analytical equations are unavailable. Qualitative Reasoning aims to predict system behavior using qualitative relationships (e.g., "if temperature increases, pressure increases"). Prior systems like GIZMO and QPE attempted "total envisionment"—generating every possible state a system could occupy. This approach led to a computational bottleneck, making them impractical for real-world problems. QPC addresses this by separating reasoning into two distinct stages: model building and simulation.
Methodology: The Hybrid Architecture
The core innovation of QPC is its two-stage process:
- Model Building (The "Compiler"): QPC uses a domain theory—a library of "model fragments" such as mass conservation or liquid flow—and a specific scenario description. Using the Algernon knowledge representation system, it infers which physical laws apply. It performs a refinement process to resolve ambiguities, conducting a case analysis to create a complete set of Qualitative Differential Equations (QDEs).
- Simulation: Once the model is compiled, it is passed to QSIM, a mathematically rigorous qualitative simulation algorithm. QSIM generates the possible trajectories of the system over time.
Key Technical Features
- Model Fragments: Quantified definitions specifying participants, conditions, and consequences.
- Algebraic Inference: QPC automatically detects implicit constraints (e.g., inferring $z=w$ if $x+y=z$ and $x+y=w$), simplifying the simulation task.
- Closed Temporal Intervals: Behaviors are treated as sequences of points and closed intervals, ensuring physical processes have well-defined boundaries.
Evaluation and Benchmarks
QPC was evaluated using three distinct domain theories:
- Heat Transfer Benchmark: A classic problem involving three objects in thermal contact with a hotter center object.
- Botany Knowledge Base: A theory of plant physiology covering water uptake and stomatal regulation.
- Nitric Acid Processing Plant: A large-scale chemical engineering domain used for hazard identification.
Key Results
QPC demonstrated significant improvements in both speed and scalability:
- Performance: On the heat transfer benchmark, QPC was 60 times faster than GIZMO, solving in approximately one minute what previously took hours and generated thousands of redundant states.
- Scalability: QPC successfully modeled a nitric acid plant with 81 active model fragments, 317 variables, and 387 constraints—the largest qualitative model solved at the time.
- Practical Utility: In botany, QPC correctly answered "what-if" questions regarding plant stress responses, matching established textbook knowledge.
Discussion and Future Work
The success of QPC highlights that decomposition is essential for qualitative modeling. By separating logic-based model building from constraint-satisfaction simulation, the system can leverage optimized algorithms for each task. However, limitations remain. QPC's reasoning about order relations (inequalities) is dependent on underlying reasoners like Bounder. Future work will focus on strengthening these reasoners and addressing higher-order continuity issues during model transitions.
Summary updated