Knowledge Level Engineering Ontological Analysis

James H. Alexander, Michael J. Freiling, Sheryl Shulman, Jeffery Staley, Steven Rehfuss, Steven MessickAAAI 1986

Introduction

In the mid-1980s, knowledge engineering was often an ad-hoc, intuitive process. Practitioners frequently struggled to transition from messy, real-world domains to structured expert systems without becoming entangled in the specifics of "if-then" rules. This foundational 1986 paper from Tektronix Laboratories addresses this challenge by proposing Ontological Analysis, a formal methodology designed to map out the "knowledge level" of a problem before implementation begins.

Drawing on Allen Newell’s concept of the knowledge level, the authors argue that analysis should focus on what a system knows about its world, independent of how that knowledge is represented in code. Without this separation, systems often become brittle, confusing, and difficult to maintain.

The Three-Step Methodology

The core of the paper is a structured approach to building an ontology, divided into three distinct layers:

  1. Static Ontology: This layer identifies the physical objects and their inherent relationships, functioning similarly to an entity-relationship model. It defines the basic "nouns" of the domain.
  2. Dynamic Ontology: This layer defines the "problem space," including the various states the system can inhabit and the operations (transformations) that change those states.
  3. Epistemic Ontology: This layer defines the knowledge—often in the form of heuristics—that guides the selection of operations at any given time.

To formalize these layers, the authors developed SPOONS (Specification Of ONtological Structure). A key variant, SUPE-SPOONS, utilizes domain equations derived from denotational semantics. These equations use mathematical operators like Cross Product ($\times$) and Discriminated Union ($+$) to define complex knowledge structures concisely.

Table 1: Static Ontology Fragment for the Intelligent Electronic Calendar (IEC).

As shown in Table 1, the static ontology defines atomic entities like Meeting and Person, and constructs complex types such as Scheduled_Meeting through cross products. This formalization ensures that the structure of the domain is mathematically sound before any rules are written.

Table 2: Dynamic Ontology for the IEC, defining states and operations.

The dynamic ontology (Table 2) further extends this by defining the State of the system as a product of meetings, purposes, and schedules, alongside the operations that transition between these states. This allows for a clear mapping of how a system moves through a problem space.

Key Insights and Domain Applications

The methodology was validated across several complex real-world "task domains," including:

  • Intelligent Electronic Calendar (IEC): Semi-automated meeting scheduling and negotiation.
  • Electronic Troubleshooting: Diagnostic logic for Tektronix oscilloscopes.
  • Medical Knowledge: A re-analysis of the MYCIN knowledge base.
  • VLSI Design: Critiquing nMOS circuitry.

Through these applications, several key insights emerged:

  • Object Classification: The authors distinguish between "first-class" objects (unique entities identified by tokens) and "second-class" objects (aggregations defined by their components).
  • Intensional vs. Extensional Entities: They emphasize separating a physical object from its various descriptions. For instance, different calendar systems might describe the same time interval using different notations (e.g., Gregorian vs. Japanese eras).
  • Rules as Associations: Rather than procedural code, rules are viewed as simple mappings between domain patterns and actions, facilitating reuse across different tasks.

Conclusion

The paper demonstrates that formal domain equations can represent highly complex logic—such as multi-party negotiation—in a way that is far more abstract and readable than raw production rules. While the authors noted limitations, such as the lack of a standard "virtual machine" for knowledge engineering and the complexity of recursive domains, the methodology significantly improved the clarity of system design. This work laid the groundwork for more rigorous, mathematically sound approaches to building intelligent systems.

Summary updated

BibTeX
@inproceedings{alexander1986knowledge,
  author = {James H. Alexander and Michael J. Freiling and Sheryl Shulman and Jeffery Staley and Steven Rehfuss and Steven Messick},
  title = {{Knowledge Level Engineering Ontological Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}

Loading related content...

Figure