The Role of Agent Diagnosis in Multi-Agent Plan Repair

The paper addresses the problem of supervising the execution of a multi-agent plan (MAP), where actions are executed concurrently by a team of cooperating agents in a partially observable environment. Because of the cooperation among agents, the failure of the action of an agent may impact also the execution of the plans assigned to other agents. The paper formalizes a distributed approach to agent diagnosis where each agent monitors its own sub-plan to detect action failures and explains these action failures in terms of faults in the functionality of the agent. The paper discusses a methodology (based on agent diagnosis) for reacting to the action failure; first, MAP is revised in order to limit the harmful effects of the detected action failure; and second, a propagation step is performed to estimate the impact of the failure and to determine the missing goals (the sub-goals no longer achievable as a consequence of the failure itself).