Multiagent plans: from individual perception to synchronized plan execution

Michael Brenner

Plans devised for execution in multiagent environments must account for the specific characteristics of such domains, namely concurrency, limited knowledge of the world, limited control of how it changes, and coordinated action execution among several agents. It has been neglected by most prior formalizations of multiagent planning that, if these characteristics are not taken into account, even a plan guaranteed to achieve some global goal will possibly not be executable for lack of synchronisation between agents. It is thus important for the agents (as a group and individually) to have plans that can be executed very flexibly and that include coordinative actions and reactions to perceptions. In this talk, we develop such a rich formalization of multiagent planning. It allows to reason about concurrent interacting actions, both qualitative and metric temporal relations between events, perceptions and communicative actions, incomplete knowledge of agents as well as the emergence of mutual belief among agents. In particular, the model ensures that valid plans embed all perceptions and coordinative actions necessary for autonomous, but synchronized execution of individual plan fragments.