Introduction
In the early 1980s, diagnostic expert systems were largely built on "empirical associations"—collections of heuristic rules linking specific symptoms to known faults. While effective for familiar problems, these systems struggled with novel devices or unanticipated failure modes. Randall Davis’s 1983 paper, Diagnosis Via Causal Reasoning, proposed a shift toward "first principles" troubleshooting. By modeling the underlying structure and behavior of a device, a system could reason about faults much like an engineer using a schematic.
Methodology: Simulation and Inference
The proposed diagnostic engine operates through a cycle of simulation and inference:
- Simulation: The system uses known inputs and models of component behavior to predict expected outputs.
- Inference: Using observed outputs, the system works backward to determine what internal signals must have been to produce those results.
- Discrepancy Detection: A fault is suspected whenever the predicted (simulated) value and the inferred value at a specific point in the circuit disagree.
To manage the vast search space of potential faults, the system employs Layered Models of Interaction. It begins with the most restrictive assumptions (e.g., only components fail, wiring is perfect). If no single component failure can explain the observed discrepancy, the system "surrenders" the current model and moves to a more complex one, such as a model that accounts for bridge faults.
Case Study: The 6-Bit Adder
The paper demonstrates these principles using a 6-bit adder circuit. In the primary case study, the device is tasked with adding 21 and 19. While the correct sum is 40, the device produces an output of 36.
Initial analysis using standard component-failure models failed to identify a single faulty gate that could explain the error. This contradiction forced the system to consider more complex interaction paths, specifically unintended connections between wires.
The Locality Principle and Paths of Interaction
A central insight of the paper is the Locality Principle. The authors argue that while a fault might appear "non-local" and chaotic in one representation (like the functional logic of a circuit), it is almost always "local" in another (like the physical layout). A bridge fault caused by a solder splash is physically local—it connects two adjacent pins—even if those pins belong to gates that are functionally distant.
By tracing Paths of Interaction, the system looks beyond the wires on a schematic. It considers physical adjacency, thermal radiation, and electromagnetic interference as potential conduits for error.
Results: Diagnosing Bridge Faults
By switching to a physical representation and a bridge fault model, the system identified that no single component could be at fault. Instead, it hypothesized a bridge between functionally distinct parts of the circuit.
The system narrowed the possibilities by checking for physical adjacency in the chip layout. It identified two plausible bridge candidates (labeled X and Y in Figure 5). Through further automated testing, the system successfully pinpointed the exact unintended connection causing the malfunction.
Conclusion
This research highlights the power of reasoning from first principles by leveraging multiple representations of a device. While the 1983 implementation faced limitations—such as manual test generation and simplified electrical models—the framework established a foundation for modern automated diagnosis. By treating models as filters and seeking locality across different domains, the system can resolve complex, non-obvious hardware failures.
Summary updated