Introduction
Qualitative Reasoning (QR) provides a framework for modeling physical systems using ordinal relationships, such as "pressure is increasing." While effective for reasoning with incomplete knowledge, QR often suffers from behavioral ambiguity. For instance, a purely qualitative model might fail to distinguish between a functioning pressure regulator and a malfunctioning one if both exhibit an increase in output pressure. This "behavioral sprawl" limits the utility of QR in practical applications like automated diagnosis.
The Q2 reasoner addresses this challenge by bridging the gap between qualitative "common sense" physics and precise numerical simulations. By incorporating incomplete quantitative data—such as numerical ranges and function envelopes—Q2 can prune impossible behaviors that a purely qualitative model would otherwise predict.
Methodology: The Q2 Algorithm
Q2 is an extension of the QSIM qualitative simulation algorithm. It propagates incomplete quantitative information through a model using several key techniques:
- Landmark Ranges: Symbolic landmarks are assigned numerical intervals (e.g., $TOP = [0.9, 0.91]$).
- Function Envelopes: For monotonic relationships where the exact function is unknown (e.g., $outflow = M^+(pressure)$), Q2 employs "envelopes"—numerically computable upper and lower bounds that constrain the function's possible values.
- Constraint Propagation: Ranges are propagated across arithmetic constraints (addition, multiplication) and quantity spaces.
- Calculus Integration: Q2 uses the Mean Value Theorem to relate the change in a state variable (like water amount) to the range of its derivative (net flow) over a specific time interval.
The core mechanism of Q2 is refinement via refutation. Rather than calculating a single precise answer, the system uses mathematical bounds to prove certain qualitative paths are impossible. If a qualitative behavior violates the provided numerical ranges, it is flagged as inconsistent and discarded.
Case Study: The Bathtub System
The authors evaluated Q2 on several models, including a standard bathtub with an open drain. In a purely qualitative simulation, multiple behaviors might be predicted (e.g., reaching equilibrium vs. overflowing). By applying quantitative ranges to landmarks like the tub's height and the faucet's inflow rate, Q2 can determine which of these outcomes is physically consistent with the given data.
Key Results and Insights
Q2 demonstrates significant power in reducing model ambiguity. In a proportional heating system model, the qualitative simulation originally predicted 18 possible behaviors; Q2 successfully eliminated all but 2 of them. Furthermore, the system enables differential diagnosis. In the case of a pressure regulator, Q2 distinguished between working and malfunctioning states by showing that their predicted output pressure ranges were non-overlapping ($[1.91, 2.98]$ vs $[3.8, 6.2]$).
The paper highlights a useful "middle ground" of knowledge. We rarely possess perfect numerical data or zero data; Q2 operates effectively within the "fuzzy" bounds typical of real-world engineering and sensor measurements. Despite the added mathematical complexity, these models remained computationally efficient, running in approximately one minute on 1988-era hardware.
Limitations and Future Work
While powerful, the initial version of Q2 applied quantitative reasoning as a post-hoc process after qualitative simulation. Future iterations could interleave these processes to prune inconsistent branches of the search tree earlier. Additionally, moving from "all-or-nothing" hard bounds to probability distributions (mean and variance) would better capture the nature of real-world sensor noise. The authors also suggest that more sophisticated inequality reasoners could further enhance the system's arithmetic capabilities.
Summary updated