Continual Multiagent Planning

Abstract

In this article, we extend the Continual Planning approach presented in the preceding article to multiagent settings. In principle, the presence of other agents increases the dynamics and uncertainty of an environment. As a result, planning in such domains becomes computationally much harder. We argue, however, that collaborative agents can overcome this complexity by proactively striving to exchange knowledge and collaborating on subproblems. The article gives an overview about the planning language MAPL that enables agents to explicitly reason about their own and others’ sensory and communicative capabilities, their beliefs and mutual beliefs, and about the necessary conditions for joint behaviour. Based on this, we describe the Continual Collaborative Planning algorithm (CCP), a distributed algorithm for autonomous agents planning and acting in multiagent worlds. We then present empirical evidence of the effectiveness of our approach in a prototypical highly-dynamic multiagent system. Finally we discuss in depth several possible applications, e.g. the use of CCP in Human-Robot Interaction and in a robot able to extend its domain knowledge on-the-fly, i.e. while acting in the environment.