Tuning a Blackboard-Based Application: A Case Study Using GBB

Daniel D. Corkill, Kevin Q. GallagherAAAI 1988

Abstract

The performance of blackboard-based AI systems often hinges more on the efficiency of the underlying database than on the problem-solving logic itself. This paper presents a case study using the Generic Blackboard Development System (GBB) to tune the Distributed Vehicle Monitoring Testbed (DVMT). By decoupling the abstract blackboard representation from its physical implementation, the authors achieved significant performance gains—reducing execution time by over 60%—simply by optimizing data storage and retrieval strategies.

Introduction

In the 1980s, the blackboard architecture emerged as a powerful paradigm for complex AI tasks such as speech recognition and vehicle tracking. In this model, independent "Knowledge Sources" (KSs) contribute to a shared data structure (the blackboard) to build and refine hypotheses. However, as applications scaled, the blackboard often became a performance bottleneck. A naive implementation using simple lists leads to $O(N)$ search complexity, causing the system to spend more time on data retrieval than on actual computation. Traditionally, overcoming this "performance wall" required extensive manual rewriting of core data structures.

The DVMT system architecture, showing the interaction between Knowledge Sources, the Planner, and the Blackboard (BB).

Methodology

The authors employed the Generic Blackboard Development System (GBB) to build and optimize the Distributed Vehicle Monitoring Testbed (DVMT), a simulator for tracking vehicles via acoustic sensors. GBB's primary innovation is the separation of the abstract blackboard view from its physical storage implementation. This allows developers to experiment with different "unit-mappings" without altering the application's logic.

The study compared several storage strategies:

  • Simple Lists: A baseline approach using basic stacks.
  • Buckets: Partitioning data dimensions (e.g., time or space) into discrete ranges.
  • Multi-dimensional Arrays: Creating a grid-like retrieval system using combinations of dimensions such as Time (t), X-coordinate (x), Y-coordinate (y), and Event-Class (c).

Retrieval was optimized through a multi-stage process involving primary retrieval, procedural filtering, and pattern-based filtering.

Key Insights

The experiments revealed several critical findings:

  1. The Intuition Gap: Initial "expert" designs for storage were often outperformed by empirically tuned configurations.
  2. Retrieval Overhead: Insertion costs were negligible (under 1% of runtime), whereas retrieval—specifically pattern-based filtering—accounted for the vast majority of overhead.
  3. Dimensionality Benefits: Increasing the dimensionality of the storage grid (e.g., indexing by Time, X, and Y simultaneously) led to dramatic reductions in the number of objects scanned during retrieval.

Results

The impact of tuning was most pronounced in complex scenarios with high noise levels. In the most challenging test case (C3), the total execution time dropped from approximately 10 hours (594 minutes) to just 3.4 hours (203 minutes). Blackboard interaction time, which initially consumed 83% of the total runtime, was reduced to 32% in the best-tuned configuration.

Performance results for various blackboard storage configurations, demonstrating the impact of dimensional tuning on total execution time.

As shown in the experimental data, configurations that utilize more dimensions (such as (c (t x y))) significantly outperform the baseline empty configuration (), which represents a simple list.

Conclusion

This case study demonstrates that significant performance improvements in AI systems can be achieved through systematic tuning of the "plumbing"—the underlying data management infrastructure. While database optimization cannot replace superior problem-solving knowledge, it ensures that the system's intelligence is not throttled by inefficient data access. Future work points toward automated tuning, where systems could dynamically reconfigure their storage representations based on real-time performance metrics.

Summary updated

BibTeX
@inproceedings{corkill1988tuning,
  author = {Daniel D. Corkill and Kevin Q. Gallagher},
  title = {{Tuning a Blackboard-Based Application: A Case Study Using GBB}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure