The Advantages of Abstract Control Knowledge in Expert System Design

William J. ClanceyAAAI 1983

Introduction

In the early development of artificial intelligence, particularly during the era of expert systems, the primary challenge was not just the acquisition of knowledge, but its organization. Systems like MYCIN demonstrated remarkable performance in medical diagnosis, yet they suffered from significant architectural limitations. William J. Clancey’s seminal 1983 research addresses these issues by proposing a fundamental shift in design: the strict separation of "control knowledge" (the strategy of reasoning) from "domain facts" (the specific knowledge of a field). This paper explores the transition from the "compiled" logic of MYCIN to the abstract, modular architecture of NEOMYCIN.

The Problem: The "Spaghetti Code" of Reasoning

Early expert systems often suffered from what Clancey identifies as a "compiled knowledge" problem. In these systems, reasoning strategies were frequently embedded directly into domain-specific rules. For instance, a rule might list its conditions in a specific order not because of a logical dependency, but to force the program to ask questions in a particular sequence.

This intertwining of strategy and fact created three primary obstacles:

  1. Maintenance Difficulties: Because the strategy was distributed across hundreds of individual rules, changing the system's diagnostic approach required a massive manual overhaul of the entire knowledge base.
  2. Opaque Explainability: When a user asked the system "Why are you asking this?", the system could only provide the specific rule it was currently evaluating. It could not explain the higher-level diagnostic strategy it was pursuing.
  3. System Brittleness: The knowledge was so specific to the task that it could not be repurposed for other domains or used as a pedagogical tool for teaching students the underlying principles of diagnosis.

Methodology: The NEOMYCIN Architecture

To solve these issues, Clancey introduced NEOMYCIN, an evolution of the MYCIN system designed to make reasoning explicit and abstract. The core methodology relies on two distinct components:

1. Abstract Metarules

Instead of domain-specific rules (e.g., "If the patient has a fever, check for infection"), NEOMYCIN utilizes abstract metarules. These rules describe the process of diagnosis without mentioning medical terms. A typical metarule might state: "If a hypothesis has a subtype that hasn't been pursued, pursue it." This represents a general strategy of refinement that applies regardless of whether one is diagnosing a patient or a malfunctioning computer.

2. Structural Relations

For abstract rules to function, the domain knowledge must be organized into a structured network. NEOMYCIN uses "structural relations" such as SUBSUMES, CAUSES, and SUBTYPE. These relations act as an API for the abstract strategy, allowing the diagnostic engine to navigate the specific facts of a domain by following the structural links between them.

Key Insights

Strategy vs. Facts

The most significant insight of this work is that "control" is a distinct form of knowledge. A human expert possesses both a library of facts and a separate set of strategies for investigating those facts. By reflecting this duality in AI architecture, the reasoning process becomes transparent and modular.

The Power of Abstraction

Stripping domain terms from control logic allows the system to provide "strategic explanations." Instead of merely citing a rule about meningitis, NEOMYCIN can explain that it is currently attempting to refine a general disorder into a more specific subtype. This makes the AI's "thought process" visible to both developers and end-users.

Indexing via Structure

Structural relations serve as the essential access paths for strategies. This architecture makes the system highly extensible; if a developer wishes to implement a new strategy—such as "check common causes before rare ones"—they simply need to define a new structural relation (COMMON-CAUSE) within the data, rather than rewriting the core engine.

Results and Impact

Clancey’s qualitative results from the NEOMYCIN project demonstrated several advantages over previous architectures:

  • Generic Utility: The system’s approximately 75 metarules constitute a generic framework for diagnosis that is largely independent of medical facts.
  • Economical Representation: By factoring out "screening" logic into abstract metarules, NEOMYCIN eliminated the need for dozens of redundant, hand-coded rules that were required in MYCIN to prevent unnecessary questioning.
  • Enhanced Pedagogy: Because the strategies are explicit, the system became a viable tool for teaching diagnostic reasoning to medical students, a task for which the original MYCIN was ill-suited.

Limitations and Future Directions

Despite its successes, Clancey noted several areas for further development. While the steps of a strategy are made explicit in NEOMYCIN, the rationale—the reason why one strategy is superior to another—often remains implicit. Additionally, the "up-front" cost of designing these abstract layers is high, making it potentially inefficient for simple, single-task systems. Finally, the author observed that domains relying on perceptual pattern matching (such as signal interpretation) may be more difficult to decompose into abstract strategies than logical, diagnostic tasks.

Conclusion

The shift toward abstract control knowledge represented a major milestone in AI design. By separating the "how" from the "what," Clancey laid the groundwork for more maintainable, explainable, and robust expert systems, moving the field away from rigid rule-based structures toward flexible, strategic reasoning engines.

Summary updated

BibTeX
@inproceedings{clancey1983advantages,
  author = {William J. Clancey},
  title = {{The Advantages of Abstract Control Knowledge in Expert System Design}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}
Figure