Overview
This research introduces a method for automatically building the "simplest adequate model" of a complex physical device. By representing components as Context-Dependent Behaviors (CDBs), the system selects models that are complex enough to explain a device's expected function but simple enough to remain computationally efficient, effectively navigating the trade-off between accuracy and speed.
The Challenge of Model Selection
In model-based reasoning, engineers face a "Goldilocks" problem: a model that is too simple fails to capture critical phenomena (like a wire's heat dissipation), while a model that is too complex (including every magnetic and kinematic property of every screw) is computationally overwhelming. Traditionally, human experts manually select which physical effects to include. This paper aims to automate this process, addressing the challenge that a single component can behave differently depending on its environment.
Methods: Context-Dependent Behaviors (CDBs)
The authors propose a framework based on CDBs. A CDB is a component model that encapsulates both its mathematical behavior (equations) and the contextual constraints under which that model is valid.
The system uses three types of context to drive selection:
- Structural Context: Physical properties and connections (e.g., "Is this component metallic?").
- Behavioral Context: The state of the system (e.g., "Is the voltage drop high enough to trigger a spark?").
- Expected Behavior: A high-level description of the device's purpose (e.g., "Temperature changes should cause the pointer to move").
The selection algorithm works in two phases. First, Identification starts with the expected behavior and pulls in necessary component models to explain the causal chain. Second, Simplification iteratively tries to replace component models with simpler "causal approximations" or removes them entirely, provided the model still explains the expected behavior.
Evaluation and Results
The system was tested using a specialized knowledge base of 25 component types and approximately 150 behavior models. The test suite included 10 electromechanical devices drawn from engineering texts.
Key results include:
- Efficiency: The program constructed models for devices with 10 to 54 components in 0.5 to 8 minutes on 1992-era hardware.
- Irrelevance Filtering: In the temperature gauge example (pictured above), the system correctly ignored the thermal interaction between the battery and the atmosphere, recognizing it as irrelevant to the primary function of measuring water temperature.
- Context Sensitivity: The system correctly identified that a coil of wire should be modeled as a heat-generating resistor in a temperature gauge, but as an electromagnet in a galvanometer.
Conclusion and Future Work
The authors demonstrate that if all model approximations are "causal" (meaning a simpler model's effects are a subset of a more complex one), the search for the simplest model becomes computationally tractable (polynomial time). By providing the "expected behavior" as a filter, the system can ignore irrelevant physical phenomena. Future work aims to address limitations in handling integration over time and the automatic setting of significance thresholds.
Summary updated