The Use of Meta-Level Control for Coordination in a Distributed Problem Solving Network

Daniel D. Corkill, Victor R. Lesser

Introduction

In distributed artificial intelligence (DAI), a primary challenge involves coordinating a network of semi-autonomous nodes to solve a global problem without relying on a central controller. Centralized control often leads to communication bottlenecks and single points of failure. However, purely local decision-making can result in redundant work, conflicting actions, or idle time. This paper, originally published in 1983, introduces "organizational structuring" as a meta-level control mechanism to achieve global coherence while maintaining local flexibility.

Problem Statement: The Coordination Challenge

Distributed sensor networks, such as those used for tracking vehicles across large geographical areas, require nodes to cooperate because no single node possesses a complete view of the environment. Coordination is essential to avoid:

  • Redundancy: Multiple nodes performing identical tasks.
  • Contradiction: Nodes pursuing mutually exclusive hypotheses.
  • Inertia: Nodes waiting for external data that may never arrive.

The authors argue that nodes must balance being "on mission" globally with reacting intelligently to local sensor data.

Proposed Architecture: Meta-Level Control

The authors propose a two-tiered coordination strategy. First, Organizational Structuring provides a high-level, relatively static strategic plan defining roles and communication paths. Second, Local Elaboration allows each node to translate this plan into specific tasks based on real-time local data.

The system extends the Hearsay-II blackboard architecture. Key components include:

  • Goal Blackboards: These track a node's intentions, allowing it to reason about what it should do next.
  • Interest Areas: These act as "knobs" for meta-level control, directing a node's attention toward specific regions or levels of abstraction.
  • Planners: Modules that trigger knowledge sources based on both local data and organizational goals.
The extended Hearsay-II architecture featuring goal blackboards and meta-level control components.

Experimental Setup: The Distraction Problem

To test their approach, the authors utilized the Distributed Vehicle Monitoring Testbed (DVMT). The task involves identifying and tracking vehicles in a 2D space using acoustic sensors. A critical component of the experiment was the introduction of a "ghost track"—acoustic noise that mimics a vehicle temporarily.

The DVMT scenario illustrating the "ghost vehicle track" used to test node distraction.

This "ghost track" serves as a distractor. If a node is too "externally-directed"—prioritizing its neighbors' hypotheses over its own local data—it may waste significant resources chasing false leads.

Key Insights and Results

The research yielded several significant findings regarding node behavior in distributed networks:

  1. The Distraction Cost: In a 4-node network, purely "externally-directed" nodes (those that strictly obey network goals) took 83–86 cycles to solve the problem. In contrast, "self-directed" nodes took only 32–33 cycles. The obedient nodes were severely hindered by the ghost track.
  2. Belief vs. Focus: The authors distinguish between accepting a neighbor's data (belief) and making that data a processing priority (focus). Effective meta-level control should influence a node's focus without necessarily overriding its local sensor beliefs.
  3. Organizational Skepticism: Robustness is improved when nodes are "skeptical" of organizational goals that contradict strong local evidence.
  4. Hierarchical Filtering: A 5-node hierarchical structure proved more resilient than a flat 4-node network. The "manager" node acted as a buffer, allowing "worker" nodes to return to productive local tasks more quickly even when the manager was momentarily distracted.

Conclusion and Future Work

While the experiments were conducted on small networks (3–5 nodes), they highlight the fundamental trade-offs between local autonomy and global coordination. Future work points toward "organizational self-design," where networks can dynamically reconfigure their roles in response to performance drops or hardware failures. The authors also envision more explicit, human-readable representations of organizational roles like "Specialist" or "Manager" on the organizational blackboard.

Summary updated

BibTeX
@inproceedings{corkill1983use,
  author = {Daniel D. Corkill and Victor R. Lesser},
  title = {{The Use of Meta-Level Control for Coordination in a Distributed Problem Solving Network}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1983},
}
Figure