Abstract
This paper introduces MISQ, a system designed to bridge the gap between observing physical system behavior and understanding the underlying qualitative differential equations (QDEs) that govern it. By automating the abduction of qualitative models, the authors transition qualitative reasoning from manual "artistry" toward automated discovery. MISQ combines constraint satisfaction with dimensional analysis and relational pathfinding to reconstruct unique physical models, even when certain variables or dimensions are missing from the input data.
Introduction
Qualitative simulation, such as the QSIM framework, provides a means to understand complex systems without requiring exact numerical parameters. However, the creation of these qualitative models has historically been a manual process. While previous tools relied on pre-defined "model fragments," MISQ addresses the challenge of building models purely from behavioral observations. It asks: "Given these observations of a system changing over time, what is the simplest set of qualitative laws that could have produced them?"
The MISQ System
The MISQ process operates in three distinct phases:
1. Data Conversion
Quantitative sensor data or hand-generated behaviors are first converted into qualitative states. This involves identifying "landmark" points such as maxima, minima, or zero crossings, and determining the direction of change (increasing, decreasing, or steady).
2. Constraint Generation and Filtering
The system generates potential combinations of variables and tests them against standard QSIM constraints, such as arithmetic addition, derivatives, and monotonic functional relationships. A key innovation in MISQ is the use of dimensional analysis as a structural filter. For instance, the system recognizes that a variable cannot be added to its own derivative because their dimensions (e.g., mass vs. mass/time) are incompatible, significantly pruning the search space.
3. Model Construction and Pathfinding
If the resulting constraints form a disconnected graph, MISQ assumes a variable might be missing. It employs relational pathfinding to search for intermediate variables that could logically connect the sub-models, effectively "inventing" necessary physical quantities like "netflow" to complete the model's logic.
Evaluation and Results
The authors evaluated MISQ on several classic physical system benchmarks, including the bathtub, the U-tube, and a nonlinear pendulum.
In the Cascaded Tanks experiment—a setup where the drain of one tank provides the inflow for the next—MISQ demonstrated its robustness. Even when the "netflow" variable for each tank was intentionally omitted from the input, the system's pathfinding successfully inferred its existence and correctly linked the inflow, outflow, and amount variables.
Furthermore, when presented with two unrelated systems (Independent Bathtubs), MISQ correctly identified them as independent sub-models rather than forcing a connection, demonstrating its ability to detect system boundaries.
Key Insights
- Uniqueness Theorem: The authors prove that for "complete" and "consistent" input behavior, MISQ is guaranteed to produce a unique, maximally constrained model.
- Dimensional Heuristics: Using units of measurement as a structural filter is a powerful method for reducing the complexity of model search.
- Abducing Missing Links: The ability to infer unmeasured variables allows MISQ to reconstruct physically sound models from incomplete datasets.
Conclusion and Future Work
MISQ represents a significant step toward the automated discovery of physical laws. While currently limited to a subset of QSIM constraints and single-state models, future work aims to expand the constraint set and handle systems that undergo state transitions (e.g., a tank overflowing). Integrating robust noise handling and leveraging libraries of known physical processes could further enhance its applicability to complex, real-world domains.
Summary updated