On-line Anticipatory Planning

We consider the problem of on-line continual planning, in which additional goals may arrive while plans for previous goals are still executing and plan quality may depend on how quickly goals are achieved. Even in domains with deterministic actions, this is a challenging problem. One common and straightforward approach is reactionary planning, in which fresh plans are synthesized when a new goal arrives. In this paper, we adapt the technique of hindsight optimization from on-line scheduling to create an anticipatory planning algorithm. Using an estimate of the goal arrival distribution, we sample possible futures and use ordinary deterministic planning techniques to estimate the value of taking each possible action. We present benchmark domains inspired by crisis response and manufacturing and show how an anticipatory approach yields a superior planner that is sensitive not only to which action should be executed, but when.