The Shifting Terminological Space: An Impediment to Evolvability

William R. Swartout, Robert NechesAAAI 1986

Abstract

In the development of expert systems, "evolvability"—the ease with which a system can be updated or expanded—remains a critical bottleneck. This paper identifies the "shifting terminological space" as a primary impediment to this goal. By analyzing the limitations of traditional rule-based architectures like MYCIN, we propose the Explainable Expert Systems (EES) framework. This approach utilizes the NIKL representation language to decouple terminological definitions from problem-solving strategies. By treating the terminological space as a first-class citizen, the framework enables automatic plan discovery and principled modularity, allowing systems to adapt to new knowledge without manual reprogramming of existing logic.

1. Introduction

During the mid-1980s, expert systems demonstrated significant power in specialized domains but were notoriously brittle. When developers attempted to modify these systems or introduce new rules, the underlying logic frequently collapsed in unexpected ways. This lack of evolvability stems from how knowledge is represented and interconnected. This paper argues that the core of the issue lies in the "terminological space"—the language and concepts used to describe the problem domain—which is often implicitly defined and inextricably linked to the rules that act upon it.

2. Impediments to Evolvability

The authors identify five major hurdles that prevent expert systems from evolving gracefully:

  1. Overly Specific Representation: Knowledge is frequently hard-coded for narrow tasks, preventing the reuse of general heuristics.
  2. Confounding Knowledge Types: Domain facts, strategic problem-solving steps, and terminological definitions are often tangled within a single rule.
  3. Impoverished Control Structures: High-level logic and iteration are often implemented as "hacks" within rule-based architectures rather than being explicitly modeled.
  4. Implicit Intent: The underlying strategy or "why" behind a rule is hidden, making it difficult for the system to explain its reasoning or for developers to modify it.
  5. The Shifting Terminological Space: In most systems, the meaning of a term is defined solely by the rules that react to it. If a rule is changed, the meaning of the term shifts for the entire system, creating a fragile "glue" that holds the knowledge base together.

3. The EES Framework and NIKL

To address these impediments, the authors introduced the Explainable Expert Systems (EES) framework. Central to EES is the use of NIKL (a descendant of KL-ONE), a knowledge representation language designed to provide independent semantics for domain concepts.

The innovation of EES lies in treating the terminological space as an independent, structured entity. This is achieved through several key mechanisms:

  • Independent Semantics: Concepts, including goals and plans, are defined in NIKL based on their internal structure (roles and restrictions). Their meaning is intrinsic and does not depend on the rules that utilize them.
  • The NIKL Classifier: An automatic reasoning tool that determines subsumption relations. The classifier can automatically recognize that a specific concept (e.g., "Locate Fault in a Spacecraft") is a specialized version of a more general concept (e.g., "Locate Fault").
  • Explicit Intent: Goals are represented as structured concepts containing a requirement description (the "what"), inputs, and outputs. Plans are similarly structured, containing a capability description and a specific method (the "how").

4. Methodology and Evaluation

The EES framework was evaluated using a network of approximately 500 concepts across two primary demonstration domains:

  1. Telemetry Diagnosis System: A system designed to diagnose faults in space telemetry hardware, such as transmitters and receivers.
  2. Program Enhancement Advisor: A tool that suggests transformations to improve the readability or efficiency of software code.

By using the NIKL classifier, the system could "discover" appropriate plans for new situations. For instance, if a developer added a specialized method for diagnosing transmitters, the classifier would automatically prefer this specific plan when a transmitter fault occurred, while maintaining a general plan for "decomposable systems" as a fallback.

5. Key Insights

The primary contribution of this work is the separation of "what" from "how." By defining goals independently of the plans that achieve them, the system remains stable even as new methods are added. This "principled modularity" ensures that the terminological "glue" remains robust. Furthermore, because the system understands the hierarchical relationships between terms, it can generate more coherent natural language explanations for its actions and assist humans in knowledge acquisition by identifying missing or inconsistent definitions.

6. Limitations and Future Work

Despite its successes, the authors noted several limitations in the 1986 implementation of NIKL:

  • Transitive Relations: The lack of support for transitivity made it difficult to represent causality or "part-of" hierarchies without manual intervention.
  • Role-Value Restrictions: The system struggled to automatically resolve complex restrictions, such as "diagnose the component that serves as the input to System X."
  • Development Overhead: Building a highly structured NIKL knowledge base requires significantly more initial effort than writing simple if-then rules, highlighting the need for more advanced knowledge acquisition tools.

7. Conclusion

By making the terminological space explicit and structured, the EES framework provides a path toward more evolvable and explainable expert systems. Decoupling terminology from strategy allows for a more modular architecture where knowledge can be added or refined without the risk of systemic collapse.

Summary updated

BibTeX
@inproceedings{swartout1986shifting,
  author = {William R. Swartout and Robert Neches},
  title = {{The Shifting Terminological Space: An Impediment to Evolvability}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure