Presented at AAAI-93, this research introduces a foundational framework for merging automated planning with model-based diagnosis. The resulting system, known as the Integrated Repair System (IRS), shifts the focus from merely identifying a fault to optimizing the entire process of restoring a device to functionality.
The Problem: Diagnosis vs. Repair
Traditional model-based diagnosis systems typically aim to pinpoint the exact faulty component. However, this "diagnosis-first" approach often overlooks two critical real-world factors:
- Misaligned Goals: The ultimate objective is usually to repair the device, not just name the broken part. In some cases, it may be more cost-effective to replace an entire module containing several components than to spend significant time and resources probing to find the specific failure point.
- Hidden and State-Dependent Costs: Most diagnostic models assume that "probes" (observations) have a fixed, uniform cost. In reality, costs are often state-dependent. For example, probing a circuit board inside a sealed appliance requires "setup" actions—such as moving the unit or removing a backplane—each with its own costs and potential side effects.
IRS addresses these issues by treating both information-gathering actions (probes) and state-altering actions (repairs) within a single mathematical and planning framework.
The IRS Architecture
The IRS framework integrates three distinct AI technologies:
- Underlying Washington Language (UWL): An action representation language that distinguishes between causal effects (changing the physical state of the world) and observational effects (gathering data without altering the state).
- UCPOP Planner: A partial-order planner used to calculate the "execution cost" of reaching the necessary state to perform a specific probe or repair.
- GDE-style Diagnosis: A model-based diagnostic engine extended with the "alibis principle." This principle allows the system to handle devices whose states change over time, such as when a part is replaced during the repair process.
The Cost Function
IRS utilizes an $n$-step lookahead cost function ($C_{total}$). It evaluates the optimal next move by calculating the cost of the current plan to achieve an operation plus the expected cost of all future diagnostic and repair steps, weighted by the probability of different outcomes.
Modeling Complex Systems: The Domestic Refrigerator
To evaluate the system, the authors utilized a detailed model of a domestic refrigerator. This model includes a wiring diagram involving a thermostat, relay, compressor, fans, and a "guardette."
This test case is particularly effective because it includes physical constraints. For instance, the system must reason about the cost of moving the refrigerator away from a wall to access the backplane before it can perform certain probes or replacements.
Strategic Flexibility and Results
One of the key insights of the IRS framework is the uniformity of action. In a planning context, there is no fundamental difference between a probe and a repair; both are actions with preconditions and costs. Some actions even serve both purposes—replacing a part provides diagnostic information if the system subsequently begins to function correctly.
As shown in the evaluation tables, IRS demonstrates "rational" strategic behavior. If the prior probability of a specific part (like the "guardette") failing is high, the system may choose to probe it even if it is expensive to reach, provided the total expected cost of repair is lower than alternative strategies. The system successfully generated complex repair plans—including disconnecting power, moving the device, and replacing parts—in approximately two minutes on a SUN SPARC station.
Conclusion and Future Work
By using a planner to estimate costs, IRS avoids counterproductive probes that cost more than the components they are intended to test. While the $n$-step lookahead is computationally intensive, the framework proves that integrating planning and diagnosis leads to more efficient real-world maintenance. Future work aims to incorporate heuristics to approximate total costs more efficiently and to account for time-dependent costs, such as the urgency of repairing a freezer before its contents melt.
Summary updated