Representation of Control Knowledge in Expert Systems

Janice S. AikinsAAAI 1980

Abstract

In the development of early expert systems, the intermingling of domain-specific inference rules with procedural control logic created significant barriers to system maintenance and transparency. This paper explores the transition from implicit control mechanisms—typified by systems like MYCIN—to explicit control representations in the CENTAUR system. By utilizing a hybrid architecture of "prototypes" (frames) and production rules, CENTAUR demonstrates that separating "what" a system knows from "how" it should apply that knowledge leads to more robust, explainable, and maintainable artificial intelligence.

Introduction: The Problem of Implicit Control

During the late 1970s, rule-based production systems were the dominant paradigm for expert systems. While systems like MYCIN proved successful in clinical domains, they suffered from a fundamental architectural flaw: implicit control. In these systems, the order in which a consultation proceeded—such as the sequence of questions asked to a clinician—was often determined by the arbitrary ordering of clauses within LISP-based production rules.

To ensure a logical flow, knowledge engineers were forced to carefully sequence these clauses, effectively turning modular domain expertise into a form of "spaghetti code." This created hidden dependencies where adding or reordering a single rule could cause the system to behave erratically or ask nonsensical questions. Consequently, these knowledge bases became fragile and difficult for anyone other than the original developer to modify or expand.

The CENTAUR Architecture

To address the limitations of flat rule-based systems, the author introduced CENTAUR, a system that employs a hybrid approach to knowledge representation. CENTAUR moves control logic out of individual rules and into explicit, high-level structures.

Prototypes and Frames

CENTAUR utilizes "prototypes," which are frame-like structures representing typical patterns or categories within a domain (e.g., a specific pulmonary disease). These prototypes serve as the organizational backbone of the system's knowledge.

Explicit Control Slots

Unlike MYCIN, where control was a side effect of rule evaluation, CENTAUR includes specific "slots" within its prototypes—such as If-Confirmed or Action—that contain explicit control logic. These slots list the tasks the system should perform in a specific context, such as "Deduce the subtype of the disease" or "Print findings." This creates a clear separation of concerns: the prototype acts as a manager directing the consultation, while the production rules act as specialized workers providing specific inferences.

Hypothesize and Match

The system operates on a "hypothesize and match" cycle. Initial data is used to suggest likely prototypes. Once a prototype is selected, the control knowledge embedded within that frame guides the subsequent search for information, allowing the system to switch its behavioral strategy based on the current hypothesis.

Evaluation: From PUFF to CENTAUR-PUFF

The efficacy of this architecture was evaluated using the PUFF system, an expert system designed to interpret pulmonary function tests. The original PUFF was a MYCIN-like system consisting of approximately 60 production rules. The author re-implemented the PUFF domain within the CENTAUR framework to compare the two approaches.

This comparative study revealed that the CENTAUR-PUFF implementation was significantly more modular. Because the interaction between rules was no longer a hidden side effect of clause ordering, changes to the knowledge base became "safer" and more predictable.

Strategic Explanations and Knowledge Acquisition

One of the most significant advantages of explicit control is the improvement in system transparency. In traditional systems, users could ask "HOW" or "WHY" to see a trace of rule evaluations. CENTAUR introduced the "CONTROL" keyword, allowing users to ask about the high-level strategy. For example, the system could explain, "I am currently trying to determine findings associated with Obstructive Airways Disease," providing a conceptual context that technical rule chains often lack.

Furthermore, the prototype structure provides a natural guide for knowledge acquisition. When adding a new disease pattern to the system, the frame structure explicitly indicates which slots (data points and control tasks) need to be defined, streamlining the process for knowledge engineers.

Conclusion and Future Directions

CENTAUR demonstrates that representing control knowledge explicitly is essential for the maturity of expert systems. While this approach requires more upfront design work—specifically in defining the hierarchy of prototypes and their associated control tasks—the long-term benefits in maintainability and user trust are substantial.

Future research may explore whether "control rules" could serve as an alternative to frame-based slots or how this architecture scales to massive, interconnected domains like general internal medicine. Nevertheless, the shift toward separating inference from strategy remains a cornerstone of clean knowledge engineering.

Summary updated

BibTeX
@inproceedings{aikins1980representation,
  author = {Janice S. Aikins},
  title = {{Representation of Control Knowledge in Expert Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure