Achieving High Quality Solutions in Distributed Agent-Based Control Systems

Recent interest in distributed agent-based control system s for dynamic, unpredictable application domains has been motivated largely by the increased responsiveness of independent entities that can react locally to changes in the operating environment. The agent-oriented paradigm for software engineering provides a basis for the construction of extremely large, complex systems, in which components can be naturally distributed across a network of heterogeneous computers, without requiring a complete analysis of their interactions. For all th eir advantages, such systems are difficult to optimize. Localization of information and computation, limited communication, and diversified goals present significant challenges to the design of systems capable of achievi ng high levels of global utility. The most common approach to optimizing agent-based systems is some sort of inter- agent exchange mechanism, usually pair-wise, in which tasks are swapped between agents, subject to the constraint that system utility, and usually the utility of each agent involved in the exchange, increases with each exchange. Such approaches tend to be either inherently and significantly sub-optimal, or inordinately complex, or both. In this paper, we present a simple optimization heuristic that has enabled an agent-based system for vehicle routing to achieve results that compare quite favorably with an established, centralized heuristic approach.