In their 1984 paper, Eva Hudlicka and Victor R. Lesser addressed a fundamental challenge in artificial intelligence: enabling complex systems to autonomously recognize and rectify strategic failures. They proposed a meta-level control framework that conceptualizes suboptimal performance as a "fault" requiring real-time diagnosis and repair.
The Challenge of Static Control
During the mid-1980s, sophisticated AI architectures, such as blackboard systems, relied heavily on numerous control parameters. These parameters governed critical functions like search depth and sensor data weighting. Typically, these settings were static, optimized by developers for "average" scenarios.
However, this approach lacked adaptability. In dynamic or atypical environments—such as when a sensor malfunctions or a task deviates from the norm—static strategies often lead to inefficient or incorrect processing. The authors argued that systems require "introspection": the ability to monitor their internal states, detect when they are stuck or pursuing noise, and adjust parameters accordingly.
The FDD Framework
The proposed solution is a meta-level control architecture based on the Fault Detection and Diagnosis (FDD) paradigm. This framework operates in three distinct stages:
- Fault Detection: This module monitors the system for "symptoms" of failure, such as an excessive number of unachieved goals, an empty agenda, or the generation of low-credibility results.
- Fault Diagnosis: Upon detecting a symptom, the diagnosis module identifies the root cause by tracing the error back through the system's logic using a System Behavior Model (SBM).
- Strategy Replanning: This final stage adjusts control parameters to steer the system back toward efficient problem-solving.
The System Behavior Model (SBM)
Central to this framework is the System Behavior Model (SBM). The SBM is a hierarchical, state-transition diagram that defines how the system should ideally function. It maps "Abstracted Objects" (e.g., goals, hypotheses) to their expected attributes and transitions.
By organizing the model into clusters of states at varying levels of abstraction, the diagnostic process can efficiently isolate problem areas without needing to analyze every low-level execution step. To pinpoint root causes, the system employs four types of reasoning:
- Backward Causal Tracing: Moving backward from a failed state to identify where the logic diverged from the model.
- Comparative Reasoning: Contrasting a "successful" execution path with a "failed" one to identify the point of divergence.
- Unknown Value Derivation: Inferring hidden internal states based on available data.
- Inconsistency Resolving: Identifying discrepancies between conflicting data sources by comparing them against the "ideal" model.
Evaluation: The Vehicle Monitoring Testbed
The authors evaluated the FDD framework using the Vehicle Monitoring Testbed (VMT), a simulation where a system interprets acoustic signals to track vehicles in a 2D space.
In a key "Fault Scenario," a vehicle moves past two sensors. Sensor B malfunctions, producing high-rated noise, while Sensor A provides accurate but lower-rated signal data. Because the system was initially tuned to trust Sensor B more, it ignored the real signal to process the noise.
The FDD system successfully detected a "highly rated goal not achieved." Through backward tracing and comparative analysis, it identified two faults: a physical malfunction in Sensor B and an inappropriately low weight assigned to Sensor A. By re-weighting the sensors, the meta-level controller enabled the system to successfully track the vehicle.
Conclusion
Hudlicka and Lesser's work broadened the definition of a "fault" to include inappropriate parameter settings. While the detection and diagnosis components were highly developed, the authors noted that future work should focus on more sophisticated replanning strategies and deeper domain models to enhance error detection. This research remains a foundational example of building self-aware, robust AI systems.
Summary updated