GBB: A Generic Blackboard Development System

Daniel D. Corkill, Kevin Q. Gallagher, Kelly MurrayAAAI 1986

Introduction

Blackboard architectures represent a powerful paradigm in artificial intelligence, where multiple independent "knowledge sources" collaborate by contributing to a shared data space to solve complex, non-deterministic problems. However, in the mid-1980s, these systems were notoriously difficult to implement. Unlike rule-based or frame-based AI, no commercial "shells" existed, forcing developers to build systems from scratch. This often resulted in architectures that were both difficult to maintain and slow in execution.

GBB (Generic Blackboard) was developed at the University of Massachusetts to address these challenges. It provides a high-performance framework designed to simplify the construction of blackboard-based AI while significantly improving runtime efficiency. By decoupling the logical structure of data from its physical storage implementation, GBB enables the creation of complex, multi-dimensional systems capable of handling thousands of objects.

System Architecture and Methodology

The core innovation of GBB lies in its "database-centric" approach. Rather than treating the blackboard as a simple shared memory space, GBB treats it as a sophisticated, structured database.

Separation of Concerns

A fundamental principle of GBB is the distinction between the Blackboard Specification and the Storage Strategy. The former defines the logical structure of the data (what it looks like), while the latter defines how that data is physically indexed and stored. This separation allows for a specialized "blackboard administrator" role. This administrator can tune the system's performance—for instance, by adjusting indexing methods—without requiring changes to the application-level code written by the "application implementer."

The GBB system architecture, illustrating the separation between blackboard specifications and implementation strategies.

Automatic Code Generation

As shown in the architecture diagram, GBB utilizes non-procedural specifications to automatically generate a tailored "database kernel" in Common Lisp. This kernel is optimized specifically for the application's unique data structures, ensuring that retrieval and insertion operations are as fast as possible.

Technical Innovations

GBB introduced several features that set it apart from previous blackboard implementations:

  1. Dimensional Spaces: The blackboard is organized into "spaces" defined by specific dimensions. These can be ordered (e.g., time, coordinates) or enumerated (e.g., object types). This spatial organization allows the system to treat AI search problems as spatial retrieval problems.
  2. Composite Units: GBB provides formal mechanisms to represent complex objects composed of multiple parts, such as a "track" consisting of several vehicle sightings. These can be queried and manipulated as single entities.
  3. Spatial Indexing (Buckets): To maintain performance when scaling to thousands of objects, GBB employs a "bucket" approach for ordered dimensions. This allows for rapid retrieval of objects within specific spatial or temporal bounds, avoiding exhaustive searches of the entire database.

Pattern Matching and Integrated Filtering

Efficiency is further enhanced through GBB's sophisticated pattern-matching language. The system supports "filter-before" and "filter-after" hooks, allowing it to discard irrelevant data at the earliest possible stage of a search. This drastically reduces computational overhead and memory allocation (often referred to as "consing" in Lisp environments).

An example of the GBB pattern-matching language used for efficient unit retrieval.

As illustrated in the code snippet above, the find-units function allows developers to search for specific vehicle tracks within defined time regions and locations, utilizing the system's underlying spatial indices for maximum speed.

Results and Impact

GBB successfully unified the disparate features of landmark AI systems, such as Hearsay-II and the Distributed Vehicle Monitoring Testbed (DVMT), into a single, generic tool. Evaluation using a reimplementation of the DVMT benchmark demonstrated that GBB could handle large-scale systems containing thousands of objects with high efficiency. The system proved that a generic shell could match or exceed the performance of custom-built systems through its highly optimized, generated database kernels.

Limitations and Future Directions

While GBB revolutionized the storage and retrieval layer of blackboard systems, the initial work focused less on the "control shell"—the logic governing how and when knowledge sources are triggered. Future work identified by the authors included:

  • Compile-time Optimizations: Further refining the code generator to speed up basic operations.
  • Distributed Processing: Extending the architecture to support blackboards distributed across multiple processing nodes for parallel AI applications.

GBB remains a foundational example of how structured database principles can be applied to AI architectures to achieve both flexibility and high performance.

Summary updated

BibTeX
@inproceedings{corkill1986gbb,
  author = {Daniel D. Corkill and Kevin Q. Gallagher and Kelly Murray},
  title = {{GBB: A Generic Blackboard Development System}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure