A Multi-Level Organization For Problem Solving Using Many, Diverse, Cooperating Sources Of Knowledge

Lee D. Erman, Victor R. Lesser

Abstract

This paper introduces the Blackboard Model, a seminal architectural framework designed to coordinate multiple, independent, and specialized "Knowledge Sources" (KSs) to solve complex, error-prone problems. Developed within the context of the Hearsay-II speech-understanding system, the architecture utilizes a shared, multi-level global data structure—the blackboard—to facilitate communication and cooperation among diverse AI agents. By employing a data-directed, hypothesize-and-test paradigm, the system integrates information across various levels of abstraction, from signal processing to conceptual meaning, offering a modular and robust approach to large-scale AI challenges.

Introduction

In the mid-1970s, artificial intelligence researchers faced a significant hurdle: "large" problems like speech recognition required the integration of vast, diverse knowledge bases. To understand a spoken sentence, a system must simultaneously process acoustic signals, phonetics, syntax, and semantics. Erman and Lesser identified three primary challenges: the inherent complexity of the problem, the uncertainty of individual "expert" modules, and the rigidity of traditional, "wired-in" software architectures.

The goal was to create an organization that allowed diverse experts to cooperate flexibly, correcting each other's mistakes in an iterative loop. This led to the development of the Blackboard Model, which decouples knowledge from control and allows for a more fluid, modular approach to problem-solving.

The Blackboard Architecture

The core of the proposed organization is the Blackboard, a global, multi-level data structure that represents the current state of the problem. The blackboard is partitioned into a hierarchy of abstraction levels. In the Hearsay-II implementation, these levels range from low-level acoustic parameters to high-level conceptual representations.

The Levels and Knowledge Sources in Hearsay-II

Knowledge is encapsulated in Knowledge Sources (KSs)—independent, self-contained modules that are experts in specific domains. Crucially, KSs do not communicate directly with one another; they are "anonymous" and interact only by reading from and writing to the blackboard.

Data-Directed Activation and Hypothesize-and-Test

Instead of a central program calling functions, the system employs data-directed activation. KSs are "triggered" when they detect specific patterns or "preconditions" on the blackboard. Once activated, a KS can modify existing hypotheses or create new ones.

The system operates via a hypothesize-and-test paradigm. A KS might post a "guess" (hypothesis) on the blackboard based on partial data. Other KSs then evaluate this hypothesis against their own expertise, either strengthening it, rejecting it, or using it as a building block for a higher-level abstraction.

Hypothesis Generation and Interaction on the Blackboard

This cooperation allows the system to handle "errorful" input. For example, a phonetic analyzer might be uncertain about a specific sound, but a higher-level lexical or syntactic KS can provide top-down constraints to resolve the ambiguity.

Implementation and Results

The architecture was validated through the Hearsay-II speech-understanding system, designed for news retrieval tasks. The implementation demonstrated several key advantages:

  • Modularity: KSs can be added or replaced without restructuring the entire system.
  • Error Tolerance: The high degree of cooperation between levels allows the system to prune incorrect hypotheses effectively.
  • Parallelism: Because KSs are independent and data-triggered, the architecture is naturally suited for asynchronous parallel processing.

The Hearsay-II system successfully integrated eight levels of representation and managed competing, overlapping partial solutions, allowing the most probable interpretation of a speech signal to emerge through iterative refinement.

Discussion and Limitations

While powerful, the Blackboard Model introduces certain challenges. In parallel environments, blackboard contention can become a bottleneck as multiple KSs attempt to access the shared data structure. Furthermore, the "precondition evaluator" must be highly efficient to avoid computational overhead. Despite these limitations, the Blackboard Model remains a foundational concept in AI, particularly in the development of multi-agent systems and complex signal-processing applications.

Summary updated

BibTeX
@inproceedings{erman1975multi,
  author = {Lee D. Erman and Victor R. Lesser},
  title = {{A Multi-Level Organization For Problem Solving Using Many, Diverse, Cooperating Sources Of Knowledge}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure