Interpreting Measurements of Physical Systems

Kenneth D. ForbusAAAI 1986

Abstract

Kenneth Forbus's 1986 paper establishes a foundational framework for Qualitative Reasoning (QR), addressing the "bridge" problem of translating raw numerical sensor data into high-level qualitative physical descriptions. By combining local data segmentation with a global "envisionment" of possible states, the system uses constraint satisfaction to reconstruct consistent timelines of physical events from ambiguous measurements.

Introduction

In the mid-1980s, AI research faced a significant challenge: while numerical simulations could predict physical values, they lacked the ability to explain the underlying causal mechanisms. Conversely, qualitative models were adept at reasoning but struggled to ingest real-world sensor data. Forbus proposed a domain-independent framework to resolve this by "reading gauges"—interpreting numerical streams through the lens of physical possibility.

The Qualitative Interpretation Framework

The proposed methodology relies on two primary components: Segmentation and the Envisionment. Raw data is first partitioned into segments where qualitative properties (such as the sign of a derivative) remain constant. These segments are then mapped against an envisionment—a pre-computed graph representing every qualitatively distinct state a system can occupy and the valid transitions between them.

Qualitative states (S1-S11) representing the envisionment for the beaker system, showing transitions between empty, liquid-filled, and boiling states.

To resolve ambiguity (e.g., a specific temperature representing multiple possible physical states), the system employs Waltz Filtering. This technique prunes interpretations that lack valid physical predecessors or successors, ensuring that the final "story" is globally consistent with the laws of physics.

Case Study: The Beaker Problem

The theory was evaluated using a simulated numerical trace of a beaker containing a mixture of water and alcohol placed on a burner. The system received 117 temperature samples over time, resulting in a stepped temperature curve as different components reached their boiling points.

The simulated temperature trace (T vs. time) used as input for the beaker problem.

Initially, each data segment had between 4 and 7 possible interpretations (P-interps). For instance, a constant temperature could signify boiling or a steady state. By applying the "Simplest Action Assumption"—which posits that adjacent data segments must correspond to adjacent states in the envisionment—the system successfully deduced the complex sequence: the alcohol boiling off, followed by the water boiling, and finally the beaker heating while empty.

Key Results and Conclusion

After four rounds of Waltz filtering, the system excluded 27 inconsistent interpretations, leaving a unique global interpretation that perfectly matched physical reality. This approach demonstrated that global context can resolve local sensor ambiguities efficiently. While the framework faces challenges with combinatorial explosion in complex systems and sensitivity to sensor noise, it remains a seminal contribution to automated physical diagnosis and real-time monitoring systems.

Summary updated

BibTeX
@inproceedings{forbus1986interpreting,
  author = {Kenneth D. Forbus},
  title = {{Interpreting Measurements of Physical Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure