Computing the Fault Tolerance of Multi-agent Deployment1

There have been tremendous advances in the last decade in the theory and implementation of massive multi-agent systems. However, one major obstacle to the wider deployment of multi-agent systems (MASs) is their capability of tolerating failures. MASs that are deployed across a network can quickly “go down” due to external factors such as power failures, network outages, malicious attacks, and other system issues. Protection against such unexpected failures that disable a node is critical if agents are to be used as the backbone for real world applications. We concern the way replication can form the basis of one tool (amongst many that are needed) to prevent a MAS from succumbing to failure. By replicating agents, we hope to improve the fault tolerance of a multi-agent system. Fault tolerance and replication techniques have been extensively studied in distributed computing systems, but much less so in the multi-agent systems domain. The faults considered in this paper are those that cause disconnection (or crash) of the nodes in the network where the MAS application resides. The fault model that we consider is one where the failure of each node in the network is represented by a probability. Given such a fault model, agents that locate on the nodes have different probabilities to be unavailable, and therefore the multi-agent system as a whole has some probability of being out of function. The idea of using replication as a fault tolerance method in our work is thus that, when facing failures, at least one copy of each agent will continue to reside on a connected, working host computer (node), so that the MAS as a whole can function as a unified application. Furthermore, in this paper, we focus on the problem of measuring the probability that a multi-agent system will tolerate the node failure. We call this probability the survivability of a MAS system.