Diagnosing Multi-agent Systems

One of the key requirements in collaborating distributed applications is that their sub-systems remain synchronized during their joint operation. With increasing deployment of software ”teams”, composed of distributed applications in complex, dynamic settings, there is an increasing need to also be able respond to failures that occur in their synchronization and coordinated operation, in particular to be able to diagnose the causes for disagreements (synchronization failures) that may occur, in order to facilitate recovery and reestablish collaboration e.g., by automated negotiations. This type of diagnosis is called social diagnosis, since it focuses on finding causes for failures to maintain designerspecified relationships between sub-systems. Naive implementations of social diagnosis processes can require significant computation and communications overhead, which prohibits them from being effective as the number of subsystems is scaled up, or the number of failures to diagnose increases. We thus seek to examine in depth the communication and computation overhead of diagnosis.