Coding approach for fault-tolerance in multiagent systems

When deploying a multiagent system in chaotic environments, some form of built-in fault-tolerance is a must. Fault-tolerance includes keeping backup copies of critical data. Existing platforms often choose to distribute data copies by replication. For most purposes, replication uses storage and bandwidth inefficiently. We investigate an information theoretic approach to choosing the most effective way to distribute this data. This approach uses erasure coding as an efficient way to avoid unneeded replication and still yield good error recovery. We propose the erasure-graph model of the multiagent society. This model describes the interconnections in the society, considering also ways the links and agents can fail. We give a partitioning schema agents can use to find the best distribution, given the knowledge of the erasure-graph for the society. Computing the partition is easily distributed over the agent society.