Introduction
Qualitative Process (QP) theory, pioneered by Kenneth Forbus, provides a framework for reasoning about physical systems using "naive physics" rather than precise differential equations. By modeling quantities and their influences, AI can simulate human-like reasoning about the world. However, pure QP theory often suffers from chronic ambiguity. For instance, if a system experiences both a positive and a negative influence simultaneously, the net result is qualitatively "unknown." This paper introduces a semi-quantitative extension to resolve such conflicts by incorporating "linguistic influence sensitivities."
Methodology: Linguistic Sensitivities and Fuzzy Logic
The core contribution is the introduction of linguistic terms to describe the magnitude of influences. Instead of binary positive ($Q+$) or negative ($Q-$) labels, the model uses fuzzy sets like "weakly positive" or "strongly negative."
- Linguistic Influence Sensitivities: Magnitudes are represented on a discrete fuzzy scale (e.g., -3 to +3).
- Fuzzy Relational Influence Algorithm: A mathematical model propagates these sensitivities through an influence graph, allowing the system to combine conflicting influences based on their relative strengths.
- Annotation Management: The system manages these sensitivities at different levels of validity—Universal, Scenario, State, and Query—ensuring that extra information is only added when necessary to resolve a specific ambiguity.
Case Study: The Industrial Reaction Vessel
The theory is evaluated using a model of a continuous-flow industrial reaction vessel. This system involves reactants entering a furnace, a heat-flow process, and a chemical reaction producing off-gas and product.
In this complex setup, standard QP theory fails to predict the direction of change for critical variables when heat input is increased. The extended model successfully resolves these ambiguities by identifying which influences dominate, such as determining that the off-gas generation rate will increase despite conflicting internal pressures.
Key Insights: Conflict Triangles and Equilibrium
A significant finding is the identification of "Conflict Triangles"—structural patterns where a single parameter affects a target through two paths. These subgraphs are the primary sources of ambiguity. By focusing sensitivity annotations on these specific areas, the model maintains computational tractability without requiring a complete numerical specification of the system.
Furthermore, the author introduces "Equilibrium Temp" as a distinct parameter rather than just a landmark value. This allows the system to reason about stability and the rate of return to equilibrium after a perturbation.
Conclusion
By bridging the gap between purely qualitative and fully numerical models, this research provides a way to reason about complex physical systems with limited data. The use of fuzzy linguistic sensitivities allows for nuanced reasoning without the high information cost of traditional simulations. Future work aims to integrate these methods with belief functions to handle both imprecision and probabilistic uncertainty.
Summary updated