Introduction
In the late 1980s, Artificial Intelligence research focused heavily on Qualitative Reasoning—the capacity for computational systems to model physical environments using qualitative signs (+, -, 0) instead of precise numerical differential equations. A fundamental technique in this domain is the qualitative resolution rule, which eliminates variables to derive global relationships from local component descriptions.
However, solving these qualitative confluences is an NP-complete problem. As the resolution rule is applied, the number of potential combinations grows exponentially, leading to a "combinatorial explosion" that renders the analysis of complex systems, such as power plants or large circuits, computationally infeasible. This paper introduces a heuristic control strategy to manage this complexity.
The Joining Rule and Consolidation
The authors propose a heuristic centered on the Joining Rule (JR), which leverages the principle of "consolidation." Consolidation involves grouping interacting components into higher-level, aggregate models to simplify the global system description.
The Joining Rule dictates that if a variable $x$ appears in exactly two confluences ($E_1$ and $E_2$), the system should prioritize eliminating $x$. This action effectively "joins" the two components into a single compound component ($E_{12}$), reducing the overall number of equations and variables without triggering an exhaustive search of the state space.
Auxiliary Simplification Rules
To further optimize the assembly process, several auxiliary rules are employed:
- Equality Rule: Substitutes variables when a direct relationship (e.g., $x = -y$) is identified.
- Ritschard’s Rule: A constrained resolution method adapted from qualitative economics.
- Single-occurrence-elimination: Discards variables that appear only once and thus do not constrain other parts of the system.
- Backtracking: Provides a fallback mechanism if the heuristic reaches a state where no further simplifications are possible.
Evaluation on Physical Systems
The effectiveness of these heuristics was tested on several real-world physical models:
- Electronic Circuits: The "CE-feedback" circuit, a standard benchmark for qualitative physics.
- Thermodynamics: A model of a pressurizer for a Pressurized Water Reactor (PWR) nuclear power plant.
- Industrial Plants: Preliminary work on large-scale industrial plant models.
Key Insights and Results
The core insight of the paper is that while qualitative resolution is mathematically NP-complete, physical systems are not "worst-case" scenarios. Engineered systems are typically modular, with components interacting through a limited number of shared variables. The Joining Rule succeeds because it mimics the hierarchical abstraction process used by human engineers.
The system successfully avoided combinatorial explosion in the tested models. For the CE-feedback loop, the combination of the Joining Rule and simplification rules derived the global "assemblage" model efficiently. Furthermore, the authors proved that for "stationary" models—where internal variables remain steady if reference variables are zero—the heuristic is mathematically sound and does not lose critical information.
Limitations
Despite its success, the Joining Rule is not a universal solver. It can fail in purely mathematical systems where every variable appears in three or more equations, though the authors note such "pathologies" are rare in physical system models. Additionally, the computational cost of specific rules like Ritschard’s Rule requires careful management, and scaling to truly massive industrial systems remains an ongoing challenge.
Summary updated