Integrating Diverse Reasoning Methods in the BB1 Blackboard Control Architecture

M. Vaughan Johnson Jr., Barbara Hayes-RothAAAI 1987

Abstract

In the development of complex autonomous systems, a single reasoning paradigm—whether hierarchical planning, opportunistic focusing, or goal-directed reasoning—often proves insufficient. This paper presents the BB1 blackboard control architecture, which unifies these diverse reasoning methods into a single, cohesive framework. By treating the selection of actions as a problem-solving task on a dedicated control blackboard, BB1 allows for dynamic, real-time adaptation of reasoning styles. We demonstrate this capability using PROTEAN, a system for modeling protein structures, specifically the lac-repressor headpiece.

Introduction

During the mid-1980s, AI systems were typically specialized for a single reasoning style. Hierarchical systems excelled at top-down planning but struggled with unexpected data. Opportunistic systems reacted well to new information but often lacked long-term direction. Goal-directed systems were efficient at working backward from objectives but were rigid in their execution.

Real-world domains, such as biochemistry, require a synthesis of these approaches. Modeling the 3D structure of a protein involves following a high-level plan while simultaneously pivoting to address specific constraints or satisfying prerequisites for high-priority actions. The BB1 architecture addresses this by providing a flexible "control plan" that evolves as the problem-solving process unfolds.

The BB1 Blackboard Control Architecture

The core of the BB1 system is the blackboard model, where independent Knowledge Sources (KSs) interact via a shared data area. The innovation lies in the separation of Domain KSs, which perform the primary task (e.g., placing a helix), and Control KSs, which manage the system's strategy.

To incorporate goal-directed reasoning, two generic control KSs were introduced:

  • Satisfy-Priority-Focus: This KS triggers when a high-priority goal lacks an immediate path to fulfillment, initiating backward reasoning to identify necessary preconditions.
  • Enable-Priority-Action: This KS activates when a high-priority action is blocked by an unsatisfied precondition, posting a new "sub-goal" to the blackboard to resolve the bottleneck.

This uniform treatment of control as problem-solving allows the system to mix reasoning methods dynamically without hard-coded transitions.

Experimental Results: The PROTEAN System

The architecture was evaluated using PROTEAN, a system designed to determine protein conformations based on biochemical constraints. A detailed trace of the system's reasoning cycles (specifically Cycles 27–31) illustrates the integration of these methods.

Figure 2: The state of the BB1 control blackboard and executable agenda at Cycle 27, showing the recommendation of an 'Anchor-Object' action.

As shown in Figure 2, the system maintains a multi-level control plan consisting of Strategy, Focus, and Heuristic levels. At Cycle 27, the system is actively managing several foci (F1, F3, F4, F5, F6) and recommends an ANCHOR-OBJECT action to progress the protein model.

Figure 3: By Cycle 30, the control plan has evolved to include eight active foci, demonstrating the system's ability to manage complex, nested reasoning goals.

By Cycle 30 (Figure 3), the complexity of the control plan has increased. The system successfully manages eight active foci simultaneously. This trace demonstrates a "dynamic pivot": when a high-priority "Yoke" action was blocked, the system automatically generated a goal-directed focus to "Anchor" the necessary components. Once the hurdle was cleared, it seamlessly returned to its original hierarchical plan.

Key Insights

  • Uniformity of Control: Deciding "what to do next" is treated as just another problem on the blackboard, enabling seamless switching between planning and reaction.
  • Semantic Flexibility: BB1 uses semantic links (e.g., cause, promote, enable) rather than strict syntax, allowing for more flexible goal-seeking than traditional backward-chaining.
  • Architectural Generality: These reasoning capabilities are built into the BB1 architecture itself, making them available to any application built on the platform without requiring domain-specific hacks.

Conclusion and Future Work

The BB1 architecture demonstrates that complex problem-solving benefits from a flexible, multi-method reasoning approach. While the computational overhead of "reasoning about reasoning" is a factor, the gains in search efficiency and problem-solving robustness are significant. Future work involves fully implementing the Satisfy-Priority-Focus KS and refining the management of large KS libraries to prevent control-loop conflicts.

Summary updated

BibTeX
@inproceedings{jr1987integrating,
  author = {M. Vaughan Johnson Jr. and Barbara Hayes-Roth},
  title = {{Integrating Diverse Reasoning Methods in the BB1 Blackboard Control Architecture}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure