Distributed Coordination of Massively Multi-Agent Systems

Coordination is a key problem in massively multi-agent systems. As applications execute on distributed computer systems, coordination mechanisms must scalably bridge the network distance between where decisions are made and where they are to be enforced. Our work on the CyberOrgs model1 addresses this challenge by encapsulating distributed multi-agent computations along with computational and communication resources they require (for carrying out the application’s functions as well as for coordinating actions of the agents) plus purchasing power represented by an amount of eCash for acquiring additional resources. Resources are dened in time and space, and are owned by cyberorgs. Resource ownership changes as a result of trade between cyberorgs. Ownership of resources coupled with an effective and scalable control structure creates a predictable resource environment for multi-agent systems and their coordination mechanisms to execute in. Particularly, the coordination mechanism can reason about the possibility of successful coordinated action based on predictable communication and processing delays. This paper presents our experience with hierarchical coordination of distributed processor resource for a system of cyberorgs internally distributed across a number of physical nodes. We demonstrate that encapsulation of network resources creates a scalable opportunity for reasoning about distributed coordinated action to support decision making. Experimental results show that the CyberOrgs based resource-aware approach scalably increases opportunities for successful coordinated distributed actions involving up to 1500 agents (in much larger systems) by reducing the delay in determining their feasibility, as well as helps avoid attempts of infeasible actions.