Intelligent Supervision of Plan Execution in Multi-Agent Systems

The paper discusses a methodology for establishing a closed control loop over the execution of a plan where concurrent actions are performed by a team of agents. The initial plan is distributed among agents and each of them performs a local control loop on the progress of the sub-plan it is responsible for. As soon as an action failure is detected, the agent invokes a Local Re-Planner in order to recovery from the failure by building a new local plan. In case such a local plan does not exist, a global re-planning phase is initiated; the global re-planner receives from the agent relevant pieces of information concerning the causes of the detected failure and the actions (performed by other agents) which directly or indirectly may fail as a consequence of such a failure (failure propagation). Model-Based Reasoning techniques are used for explicitly modeling the actions evolutions both in the nominal and in the abnormal situations. These extended action models play a critical role in re-planning, since the (possibly abnormal) health states of the agents can be taken into account for predicting the effects of actions.