Abstract: This paper, originally published in 1983, introduces Procedural Expert Systems (PES), a framework designed to explicitly represent expert procedures using Recursive Transition Networks (RTNs). By combining declarative invocation with procedural bodies, PES enables modular, explainable, and opportunistic reasoning, addressing the limitations of traditional rule-based systems in handling sequential tasks.
Introduction
In the early 1980s, the field of Artificial Intelligence was dominated by "Expert Systems" that primarily utilized declarative facts. While these systems excelled at diagnostic tasks based on "what" is known, they struggled significantly with "how" tasks—the step-by-step procedures experts follow to solve problems. This paper represents a pivotal shift toward encoding procedural expertise directly into machine logic.
The Procedural Gap in Rule-Based AI
Most expert systems of the era, such as the renowned MYCIN, were strictly rule-based. They relied on vast sets of "if-then" statements. While effective for simple diagnostics, this architecture was cumbersome for representing sequences of actions. To simulate a procedure, developers often had to use "tricks," such as carefully ordering clauses within a rule to force a specific execution path.
This approach created fragile knowledge bases that were difficult to update and nearly impossible for the system to explain. The authors sought a method to make these procedures explicit without sacrificing the modularity and flexibility that made expert systems valuable.
The Knowledge Area (KA) Architecture
The core innovation of the paper is the Knowledge Area (KA). A KA is a modular unit of expertise consisting of two distinct parts:
- Invocation Part: A logical expression defining when the knowledge is applicable. This allows for two types of triggers: goal-directed (triggered when the system needs to find a specific fact) and data-driven (triggered when new data is observed).
- Body: A Recursive Transition Network (RTN). Unlike a flat list of rules, the body is a map of states and transitions. Arcs between states are labeled with tests (to validate a path) and actions (to conclude facts or perform tasks).
By using RTNs, the "how-to" of an expert is represented naturally. This structure allows the system to manage a Goal Stack, where multiple KAs can be relevant simultaneously, and the system can execute them nondeterministically.
Implementation and Evaluation: The Peritus System
To test their framework, the authors developed Peritus, a tool that compiled these transition networks into LISP code for efficient execution. They applied Peritus to the domain of automobile engine fault diagnosis, using expert information from automotive service manuals.
The system demonstrated a human-like testing sequence. For instance, in a sample consultation, the system guided a user through electrical diagnostics, correctly identifying a faulty ignition coil with a certainty factor (CF) of 0.8 and a faulty coil lead with a CF of 0.5. This was achieved by following a logical progression: checking the spark, then the coil, then the distributor.
Key Insights: Opportunism and Explanation
- Opportunistic Reasoning: Purely goal-directed systems are often too rigid. PES allows a data-invoked KA to "wake up" and interrupt the current process if a significant fact is observed. If a mechanic checking the fuel system notices a specific engine knock, the system can pivot immediately to investigate a head gasket failure.
- Improved Explanations: Because the procedure is explicit, the system can explain its reasoning by pointing to its current state in the transition network. This provides significantly better context to the user than a simple chain of "if-then" rules.
- Judgmental Knowledge: The authors integrated certainty factors to handle the "messiness" of real-world data, allowing the system to reason with probabilities rather than binary true/false values.
Limitations and Legacy
The authors noted that allowing arbitrary logical expressions in the invocation part could lead to performance bottlenecks and complexity in meta-reasoning. Additionally, managing "used" KAs—deciding when a partially completed procedure should be discarded—remained a technical challenge.
Despite these hurdles, this work laid the groundwork for Procedural Reasoning Systems (PRS), which continue to influence robotics and autonomous agent research today. It was an early, successful attempt to provide AI with the "common sense" of a sequence-oriented human expert.
Summary updated