Reactive Planning with Uncertainty of a Plan

An autonomous agent in the real world needs to do planning depending on the dynamic world. For making a planning system adaptive to the dynamic world, various methods on reactive planning have been proposed. Interleaving planning with execution is a general and natural approach. However, it has a significant problem: when planning should be switched to execution, and few solutions has been proposed. In this paper, we propose a theoretical framework in which the switching timing is determined with the success probability. The success probability represents uncertainty in the execution of a plan, and depends on a change of the real world. Furthermore, since the success probability is obtained based on causality between literals, it is smciently precise. In our method, planning is switched to executionlobservation only when the success probability decreases less than thresholds. Since the success probability is used as an evaluation function,for searching a plan, we present a new interleave planning algorithm with the success probability, and different experimental results from conventional planning are obtained.