Designing a Network Infrastructure for Survivability of Multi-Agent Systems

In this paper we consider a society of agents whose interactions are known. Our objective is to solve a strategic network infrastructure design problem to determine: (i) number of nodes (usually computers or servers) and their processing speeds, (ii) set of links between nodes and their bandwidths, and (iii) assignment of agents to nodes. From a performance standpoint, on one hand all the agents can reside in a single node thereby stressing the processor, on the other hand the agents can be distributed so that there is a maximum of one agent per node thereby increasing communication cost. From a robustness standpoint since links and arcs can fail (possibly due to attacks) we would like to build a network that is least disruptive to the multiagent system functionality. Although we do not explicitly consider tactical issues such as moving agents to different nodes upon failure, we would like to design an infrastructure that facilitates such agent migrations. We formulate and solve a mathematical program for the network infrastructure design problem by minimizing a cost function subject to satisfying quality of service (QoS) as well as robustness requirements. We test our methodology on Cougaar multi-agent societies.