Device Representation and Reasoning with Affective Relations

This paper introduces "affective relations" as a novel way to represent complex devices for diagnostic reasoning. This approach bridges the gap between traditional heuristic expert systems and more complex model-based reasoning, making advanced diagnostics feasible for systems where full models are impractical to build.

TL;DR

This paper proposes "affective relations" to represent complex devices for diagnostic reasoning, offering a middle ground between heuristic expert systems and full model-based approaches. This method allows for more robust and maintainable diagnostic systems by capturing expert knowledge at a higher, more abstract level, particularly for self-checking systems.

What Problem Does This Paper Address?

The paper addresses the limitations of traditional "first-generation" expert systems, which suffer from issues like rule complexity, difficulty in modification, and implicit knowledge. While "second-generation" model-based approaches offer more principled reasoning, they often require extensive and costly development of detailed device models. This is particularly challenging for highly complex systems like modern telephone switching equipment, where creating a complete, accurate model is often infeasible. The authors aim to provide a practical solution that captures the benefits of model-based reasoning without the prohibitive modeling cost, especially for systems that already possess a degree of self-checking capability.

Methods

The core innovation is the concept of "affective relations," which define a high-level, abstract, non-behavioral representation of devices. Instead of modeling detailed component behavior, affective relations capture diagnostically motivated connections between components. These relations are designed to be easily acquired from domain experts, mirroring how they intuitively link symptoms to potential causes.

Key elements of their device model include:

  • Classes: Basic types like Device and Alarm.
  • Affective Relations: immediate-part-of, part-of, subpart, depends-on, dependent, standby. These describe structural and functional dependencies.
  • Other Relations: self-alarm, part-alarm, dependent-alarm, functional, simplex. These link devices to alarms and their operational states.
  • Properties and General Relationships: These define the semantics of the relations, enabling inference (e.g., transitivity of part-of, symmetry of standby, inverse relationships).

Diagnostic reasoning is performed by applying general inference rules to this affective relation model. This approach replaces ad-hoc heuristics with principles derived from the device's structure and dependencies. The system dynamically builds and refines the device model based on incoming alarm messages.

Datasets

The paper uses the 4ESS® telephone switching system as a case study. The authors describe a "rational reconstruction" of an existing heuristic expert system for monitoring this system, demonstrating how their affective relation approach can replace its rules. No new datasets are introduced; rather, the focus is on modeling an existing complex system.

Key Insights and Ideas

  • Bridging the Gap: Affective relations offer a practical middle ground between brittle heuristic systems and overly complex model-based systems, making advanced reasoning accessible for real-world, complex devices.
  • Expert Knowledge Abstraction: The approach effectively captures the abstract, high-level knowledge that human experts use for diagnosis, which is often easier to acquire than detailed behavioral models.
  • Dynamic Model Building: The system dynamically constructs and refines the device model as alarms arrive, making it adaptable to variations in system configurations.
  • Generality and Clarity: Replacing heuristics with general diagnostic rules based on affective relations leads to more general, clearer, and more maintainable diagnostic systems.
  • Self-Checking Systems are Key: The approach is particularly well-suited for complex, self-checking systems where alarms are already generated, as it focuses on interpreting these alarms within a structured model.

Key Results

The authors report that their implemented reconstruction of the 4ESS alarm monitor achieved approximately 10% of the original expert system's functionality with a significantly simpler and more organized knowledge base. They believe their methodology can be extended to cover the entire 4ESS monitoring problem. The key qualitative result is the demonstration that diagnostic rules based on affective relations can replace complex heuristics, leading to more general and robust reasoning. For instance, they show how specific heuristics for alarm correlation, handling simplex mode, and suspecting clocks can be replaced by more general principles like "If a device's standby is not functional, then alert on any alarm" or "Concurrent alarms on multiple dependents of a device indicate a problem with that device."

Limitations and Future Work

  • Current Functionality: The implemented prototype covers only a portion of the original system's functionality.
  • Shared Ontology: A future objective is to develop a shared ontology that can be used to monitor and diagnose many different kinds of devices, raising the question of how general causality and diagnostic principles are across device types.
  • Non-Self-Checking Systems: The authors suggest exploring how their approach could be adapted for systems that are not inherently self-checking, by focusing on reasoning at the appropriate level of abstraction.
  • Pragmatic Solutions: A driving force is to develop pragmatic solutions that balance limited developer resources with the desire for the benefits of "ideal" diagnostic approaches.

Summary updated

BibTeX
@inproceedings{1995device,
  title = {{Device Representation and Reasoning with Affective Relations}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}

Loading related content...

Figure