Abstract
The development of blackboard architectures in artificial intelligence has historically been hampered by a trade-off between system flexibility and computational efficiency. This paper introduces the Generic Blackboard development system (GBB), which utilizes data abstraction and n-dimensional spatial indexing to decouple application logic from database implementation. By separating the specification of blackboard objects from their physical storage structures, GBB enables the creation of high-performance AI systems that remain adaptable to evolving research requirements. The framework is evaluated through its application in complex domains such as speech recognition and vehicle monitoring.
Introduction
In early AI research, the blackboard architecture emerged as a powerful paradigm for collaborative problem-solving among independent "Knowledge Sources" (KSs). However, developers often faced a critical dilemma: choosing between slow, general-purpose databases that offered flexibility, or fast, hand-coded databases that were brittle and difficult to modify. This paper from the University of Massachusetts Amherst addresses this tension by introducing the Generic Blackboard development system (GBB), a framework that allows researchers to achieve both high performance and architectural flexibility through advanced data abstraction.
The Blackboard Dilemma
In a blackboard system, multiple KSs collaborate by reading from and writing to a shared global database. Historically, three approaches dominated:
- Unstructured Lists: Flexible but computationally expensive as the object count grows.
- General-Purpose Kernels: Utilizing relational databases, which provide flexibility but incur massive overhead.
- Customized Kernels: Hand-coded for specific tasks to maximize speed, but requiring total rewrites of KSs whenever data structures changed.
The authors argue that for complex tasks like vehicle tracking, the primary bottleneck is the "interaction/computation ratio"—the time spent finding data versus processing it.
The GBB Methodology
GBB resolves these issues through a three-phase development methodology that separates the logical view of data from its physical implementation:
- Blackboard & Object Specification: Developers define n-dimensional "spaces" (e.g., time, x-position, y-position) and the "units" (objects) that inhabit them.
- Application Coding: Knowledge Sources are written using abstract, declarative patterns. Instead of direct memory addressing, a KS requests data based on spatial or temporal criteria.
- Implementation Specification: Separately from the logic, the developer defines how data is stored (e.g., as a 2D array or linked list).
As illustrated in the figure above, the creation of a hypothesis (make-hyp) is handled through a keyword-based interface. This allows the system to manage attributes like :TIME-LOCATION-LIST and :EVENT-CLASS without the developer needing to know the underlying storage mechanics. GBB’s code generator then "knits" these specifications into a highly efficient, customized executable.
Key Insights: N-Dimensional Abstraction
By viewing the blackboard as a metric space, retrieval becomes a geometric problem. GBB allows for efficient "associative retrieval" by simulating it in software, hiding the actual "locators" (pointers or indices) behind an abstraction layer. This information hiding ensures that the application logic remains robust even if the blackboard structure is reorganized or the underlying storage method is tuned for performance.
Evaluation and Results
The GBB framework was evaluated using two landmark AI systems: Hearsay-II (speech understanding) and the Distributed Vehicle Monitoring Testbed (DVMT). The results demonstrated that GBB could provide the performance of a hand-customized system while maintaining the ease of use of a general-purpose tool. Developers could "tune" the blackboard representation—for instance, switching from a vector to a multi-dimensional array for indexing—to match data density without changing a single line of Knowledge Source code.
Conclusion
GBB represents a significant advancement in blackboard architecture design by proving that data abstraction can effectively decouple AI logic from database implementation. While the initial setup requires a more complex conceptual framework, the resulting systems are both highly efficient and remarkably flexible, providing a scalable foundation for complex, real-time AI applications.
Summary updated