Community-based load balancing for massively multi-agent systems

Recently, large-scale distributed multiagent systems consisting of one million of agents have been developed. When agents are distributed among multiple servers, both the computational and interaction cost of servers must be considered when optimizing the performance of the entire system. Multiagent systems reflect the structure of social communities and artificial networks such as the Internet. Since the networks possess characteristics common to the 'small world' phenomenon, networks of agents on the systems can be considered as small worlds. In that case, communities, which are the sets of agents that frequently interact with each other, exist in the network. Most previous works evaluate agents one by one to select the most appropriate agent to be moved to a different server. If the networks of agents are highly clustered, previous works divide the communities when moving agents. Since agents in the same community often interact with each other, this division of communities increases the interaction cost among servers. We propose community-based load balancing (CLB), which evaluates the communities to select the most appropriate set of agents to be moved. We conducted simulations to evaluate our proposed method according to the network of agents. Our simulations show that when the clustering coefficient is close to 1.0, the interaction cost with CLB can be approximately 30% lower than that with previous works.