HEARSAY-III: A Domain-Independent Framework for Expert Systems

AAAI 1980

TL;DR

This paper introduces HEARSAY-III, a domain-independent framework for building expert systems that generalizes the successful HEARSAY-II speech understanding system. It addresses the challenge of creating flexible and adaptable expert systems by separating domain knowledge from control knowledge and introducing a novel scheduling blackboard mechanism.

What Problem Does This Paper Address?

Expert systems, particularly those based on production rules, have been a cornerstone of AI research. However, building these systems often involves tightly coupling domain-specific knowledge with the control mechanisms that dictate how that knowledge is applied. This makes systems rigid and difficult to adapt to new domains or modify. The HEARSAY-II system demonstrated the power of a blackboard architecture for managing complex, incremental knowledge processing, but it was domain-specific. The authors aim to generalize this approach, creating a framework that allows for the development of expert systems that are easier to build, more adaptable, and where the scheduling (control) itself can be managed by expert knowledge.

Methods

HEARSAY-III builds upon the HEARSAY-II architecture by introducing two key generalizations:

  • Domain Independence: The core framework is designed to be independent of any specific domain. This is achieved by generalizing the blackboard structure and the way knowledge sources (KSs) interact.
  • Expert Scheduling: Recognizing that scheduling (deciding which KS to execute next) is a complex task, HEARSAY-III treats scheduling itself as a problem that can be solved by expert knowledge. This is implemented using a scheduling blackboard.

Key methodological components include:

  • Blackboard Structure: Generalized from HEARSAY-II's fixed levels to a more flexible, hierarchical class structure. Units (replacing hypotheses) are instances of classes, allowing for inheritance and custom domain structures. A Choice Set mechanism is introduced to represent and manage alternative interpretations at decision points.
  • Scheduling Blackboard: A separate blackboard dedicated to managing the execution of KSs. Activation records for KSs are units on this blackboard. Scheduling KSs can interact with this blackboard to prioritize, organize, and select which domain KSs to execute.
  • Context Mechanism: A system for managing alternative lines of reasoning, similar to those found in older AI languages like QA4 and CONNIVER. This allows for controlled exploration of different interpretations and problem-solving paths.
  • Relational Database (AP3): The blackboard and all data structures are implemented using a general-purpose, typed, relational database system. KS triggering patterns are translated into AP3 demons.

Datasets

Not discussed in the paper. The paper focuses on the framework's architecture and capabilities rather than specific empirical evaluations on large datasets.

Key Insights and Ideas

  • Separating Competence and Performance: The framework explicitly separates "competence knowledge" (domain-specific KSs that generate legal moves/interpretations) from "performance knowledge" (scheduling KSs that control search). This separation is crucial for building adaptable systems.
  • Scheduling as an Expert Problem: Treating scheduling as a problem solvable by expert knowledge (scheduling KSs) is a powerful idea. It allows the system's control strategy to be as flexible and adaptable as its domain knowledge.
  • Generalizing the Blackboard: Moving from fixed levels to a class-based hierarchy for the blackboard makes the architecture significantly more flexible and applicable to a wider range of domains.
  • Choice Sets for Decision Points: The Choice Set mechanism provides a clean way to represent and manage points where multiple interpretations or actions are possible, allowing for explicit representation of unmade decisions.
  • Contexts for Controlled Search: The context mechanism offers a robust way to manage and explore alternative reasoning paths without interfering with the main line of inquiry.

Key Results

The paper describes the architecture and implementation of HEARSAY-III. It mentions testing on two small applications (a cryptarithmetic problem and a cryptogram decoding problem) and ongoing major implementation efforts for the SAFE system, a natural language description generator, and a program transformation system. The authors state that the architecture "seems to be a helpful one" and that preliminary experience indicates performance (scheduling) is a difficult but manageable issue.

Limitations and Future Work

  • Performance: The authors acknowledge that performance, particularly scheduling, is a difficult issue. While the architecture provides flexibility for developing scheduling algorithms, achieving efficient performance remains a challenge.
  • Implementation Effort: While the initial design was straightforward due to leveraging AP3, efficiency considerations have necessitated substantial implementation effort.
  • Future Work: The paper implies that further work will focus on refining the scheduling mechanisms and applying the framework to more complex applications. The flexibility of the architecture is seen as a key enabler for simplifying the development of sophisticated scheduling algorithms.

Summary updated

BibTeX
@inproceedings{1980hearsay,
  title = {{HEARSAY-III: A Domain-Independent Framework for Expert Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure