A Distributed Communication Architecture for Dynamic Multiagent Systems

We investigate the problem of creating a robust, rapidly converging, Distributed Communication Architecture (DCA) for the domain of low bandwidth, single channel Multiagent Systems (MAS) in which agents may drop in and out of communication without prior notification. There are only three capability-based assumptions made by the algorithm: 1) agents can classify a signal's message as either noise, silence, or clarity, 2) agents can classify their own messages, and 3) agents can understand one another to some degree. The final structure allows agents to communicate in a round-robin manner without any centralized or hierarchical control. We evaluate DCA's the convergence rate through four distinct experiments, including both a worst-case scenario that consists of all agents starting simultaneously and a more common-case scenario in which agents offset their starting times. We examine effective throughput as the average number of clearly sent messages in a cycle to determine the amount of information successfully communicated. Lastly, we emulate situations found in problems with moving agents to show that agents who incorporate local observations can improve both their convergence rates and throughput.