Pheromone Learning for Self-Organizing Agents

Abstract

A central issue in distributed systems engineering is enabling agents with only a local view of their environment to take actions that advance global system objectives. One example of this tension is that individual agents may take actions that consume system resources, even when they are not advancing the overall system objectives. Thus, paradoxically, system performance can sometimes improve if individual agents reduce their activity. Agents in such systems need a way to modulate their individual behavior in the light of the system's state, preferably in a way that does not require centralized control. We illustrate the problem of hyperactive agents in three application domains. We describe a simple, decentralized scheme, inspired by insect pheromones, that enables individual agents to adjust their level of activity as the system operates, and extend this mechanism to provide a general approach for dealing with approaching deadlines. Then, we demonstrate the effectiveness of these mechanisms in the example domains.