An Organizational Approach to Adaptive Production Systems

Toru Ishida, Makoto Yokoo, Les GasserAAAI 1990

Abstract

Traditional parallel production systems often face a trade-off between real-time performance and resource efficiency. This paper proposes an "Organization Self-Design" (OSD) approach for distributed production systems. By allowing agents to autonomously decompose (split) when overloaded and compose (merge) when idle, the system dynamically adapts its parallelism to meet real-time deadlines while minimizing hardware waste.

Introduction

Distributed artificial intelligence (DAI) systems must often handle unpredictable, real-time workloads. Rigid, monolithic architectures struggle in these environments. If resources are under-allocated, the system misses deadlines during peak loads; if over-allocated for the worst case, hardware is wasted during idle periods. This research explores a middle ground: a system that reorganizes its own structure on the fly to match current demand.

Distributed Production Systems and Agent Architecture

The authors transitioned from a parallel production system with global control to a Distributed Production System with distributed control. In this model, each agent is a self-contained unit with its own production memory (rules), working memory (facts), and "organizational knowledge."

Figure 4: Distributed agent architecture showing boundaries of organizational knowledge and rule dependencies.

As shown in Figure 4, an agent (e.g., agentP) maintains knowledge about its dependencies and potential interferences with other agents. This localized knowledge allows for synchronization without a global controller, using a "deactivate/activate" messaging protocol to prevent conflicts when multiple agents fire rules simultaneously.

Organization Self-Design (OSD)

The core of the adaptive mechanism lies in two primitives:

  • Decomposition: When an agent’s load (firing ratio) exceeds a threshold (e.g., 80%), it creates a new agent and transfers half of its rules and data, increasing parallelism.
  • Composition: When the organization-wide load is low and an agent’s firing ratio drops below a threshold (e.g., 30%), it merges with a neighbor to free up resources.

Reorganization occurs incrementally at the end of each problem-solving cycle, allowing the system to evolve gradually without halting.

Experimental Evaluation

The approach was evaluated using a simulation of the Waltz labeling problem, a classic AI benchmark. The simulation compared the adaptive OSD approach against a conventional parallel approach with a fixed number of agents.

Figure 5: Performance comparison showing response time and number of agents over time for (a) constant load and (b) periodic load.

Key Results

The experiments yielded several critical insights:

  • Dynamic Stability: Under constant request intervals, the system reached a stable state where workload was equalized across agents (Figure 5a).
  • Resource Efficiency: In periodic load tests (Figure 5b), the OSD approach dynamically scaled between 4 and 17 agents. While a static system would require 17 agents to meet peak demand, the OSD approach averaged only 8.95 agents—a nearly 50% saving in resource utilization.
  • Deadline Adherence: The adaptive approach was significantly more effective at meeting response-time limits compared to a fixed-size organization with the same average number of agents.

Conclusion and Future Work

The OSD approach demonstrates that the structure of an agent society is a powerful lever for adaptation in real-time AI. By treating parallelism as a managed resource, systems can balance performance and efficiency. Future work includes addressing the "lag" in reorganization during sudden spikes, refining heuristic thresholds, and testing on actual message-passing multiprocessor hardware to account for real-world communication overhead.

Summary updated

BibTeX
@inproceedings{ishida1990organizational,
  author = {Toru Ishida and Makoto Yokoo and Les Gasser},
  title = {{An Organizational Approach to Adaptive Production Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}
Figure