StarPlan II: Evolution of an Expert System

Ronald Siemmens, Marilyn Golden, Jay C. FergusonAAAI 1986

Presented at AAAI-86, this research chronicles the transition of StarPlan, an expert system designed for satellite anomaly resolution, from a traditional rule-based architecture to a more robust, model-based approach. By moving away from "spaghetti" rule sets and toward a structured, causal reasoning framework called PARAGON, the authors demonstrate how to build diagnostic systems that are more maintainable, modular, and capable of handling complex, multi-fault scenarios.

The Problem: The Limits of Rule-Based Systems

Managing a satellite in real-time is a high-stakes task involving massive streams of telemetry data. When something goes wrong, operators must quickly perform monitoring, situation assessment, diagnosis, and planning.

The first generation, StarPlan I, used a "Society of Experts" approach—a collection of specialized rule-based modules. However, as the system grew, it hit several walls:

  1. Complexity & Maintenance: Rule-based systems become "brittle" and difficult to test because a change in one rule can have unpredictable side effects elsewhere.
  2. Multiple Faults: When a single failure causes symptoms across different subsystems, the independent "experts" struggled to coordinate, leading to a combinatorial explosion of necessary rules.
  3. Lack of Semantics: Rules often captured what to do but not why the system behaved that way, making it impossible to use generic reasoning algorithms.
The original StarPlan I architecture utilized a 'Society of Experts' approach.

The Evolution: StarPlan II and PARAGON

StarPlan II fundamentally restructures how knowledge is represented and used. The core innovation is the PARAGON knowledge representation scheme. Instead of relying solely on production rules, it uses a hybrid approach combining object-oriented frames and declarative behavioral descriptions.

Functional Decoupling

StarPlan II breaks the diagnostic process into five distinct modules, creating a more streamlined and logical flow compared to its predecessor:

  1. Active DataBase: Monitors telemetry and flags when data violates expectations.
  2. Situation Assessment: Focuses the system's attention on the most relevant objects involved in an event.
  3. Causal Diagnosis: Uses the object models to trace failures back to their root cause.
  4. Goal Determination: Identifies the desired "safe state" for the satellite.
  5. Planning & Command: Generates the specific sequence of commands to reach that goal while respecting operational constraints.
The StarPlan II architecture features a modular, functionally decoupled design.

Key Insights and Results

  • Heuristics vs. Models: The authors argue that while rules are good for "rules of thumb" (heuristics), they are a poor way to describe physical systems. Modeling the causal relationships between components is far more powerful for diagnosis.
  • Generic Reasoning: Because the knowledge is structured semantically in PARAGON, the diagnostic algorithms themselves become generic. You don't need to write new code for a new satellite; you just provide a new object model.
  • Separation of Concerns: By separating "what is happening" (Situation Assessment) from "why it happened" (Diagnosis), the system avoids the coordination headaches found in StarPlan I.

Conclusion

The transition to StarPlan II successfully separated classification, diagnosis, and planning, reducing the overlap and complexity of the knowledge base. While production rules are still necessary for areas where information is not well-defined enough to be modeled, the model-based approach offers a significant leap in robustness and maintainability for aerospace expert systems.

Summary updated

BibTeX
@inproceedings{siemmens1986starplan,
  author = {Ronald Siemmens and Marilyn Golden and Jay C. Ferguson},
  title = {{StarPlan II: Evolution of an Expert System}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}

Loading related content...

Figure