Abstract
Distributed Artificial Intelligence (DAI) systems often struggle with static architectures in dynamic environments. This paper presents a framework for Organization Self-Design (OSD), enabling multi-agent systems to autonomously restructure themselves. By employing primitives for agent decomposition and composition based on local and global performance metrics, the system maintains an optimal balance between parallel processing and communication overhead.
Introduction
In early distributed AI, system architectures were typically fixed at design time. A system with a set number of agents would remain static regardless of fluctuations in workload or environmental demands. This creates a "correspondence" problem: for maximum efficiency, a system's internal structure must align with the problem's structure and environmental constraints. When tasks arrive at variable rates or with strict deadlines, static architectures often become bottlenecks. The authors propose moving away from external intervention toward agents equipped with the "social" knowledge required for autonomous reorganization.
Methods: The SDPSA Architecture
The authors introduce the Self-Organizable Distributed Production-System-Based Agent (SDPSA). This microarchitecture allows agents to interleave reorganization logic with domain-specific problem-solving.
Reorganization Primitives
The core of the OSD framework relies on two primary operations:
- Decomposition: Splitting a heavily loaded agent into two to increase parallelism.
- Composition: Merging two adjacent agents to reduce inter-agent communication costs.
Decision Logic
Agents monitor two key metrics to trigger these primitives:
- Local Activity Ratio ($R$): A measure of how busy an individual agent is.
- Organizational Performance ($PERF_{required}$): A global requirement indicating whether the system as a whole is meeting its time-limit constraints.
Reorganization rules are treated as high-priority production rules, ensuring that structural changes are addressed during the agent's "select" phase before standard problem-solving tasks.
Evaluation and Results
The architecture was tested using a simulation of the Waltz labeling program, a classic AI constraint-satisfaction task for interpreting 3D line drawings. The simulation involved 36 rules and required 80 firings per solution.
Performance and Efficiency
The self-designing system achieved performance parity with systems requiring human-triggered reorganization. Notably, the OSD approach proved significantly more resource-efficient. While a traditional static parallel system might require 17 agents to meet specific performance demands, the self-organizing system met the same requirements using an average of only 9 agents.
Stability and Overhead
A critical finding was the impact of Reorganization Overhead ($O_r$). Structural changes consume time during which an agent cannot solve domain problems. If this "blackout" period is too long (exceeding 30 cycles in the simulation), it can lead to "oscillations"---cascades of unnecessary merges and splits caused by agents misinterpreting their neighbors' temporary inactivity.
Conclusion
This work conceptualizes agents not as fixed entities, but as fluid boundaries within a "fabric of knowledge." By allowing these boundaries to shift dynamically, systems can adapt to environmental pressure. Future work aims to move beyond heuristic rules toward analytic cost-benefit models for reorganization and implementing damping mechanisms to prevent structural instability under high overhead.
Summary updated