|
Efficient and effective distributed planning requires careful control over how much information the planning agents broadcast to one another. Sending too little information could result in incorrect plans, while sending too much information could overtax the distributed planning system’s resources (bandwidth and computational power). Ideally, distributed planning systems would have an efficient technique for filtering a large amount of irrelevant information from the message stream while retaining all the relevant messages. This paper describes an approach to controlling information distribution among planning agents using {\em irrelevance reasoning} \cite{LevySagiv93}. In this approach, each planning agent maintains a data structure encoding the planning effects that could potentially be relevant to each of the other agents, and uses this structure to decide which of the planning effects that it generates will be sent to other agents. We describe an implementation of this approach within a distributed version of the SIPE-2 planner. Our experiments with this implementation show two important benefits of the approach: first, a noticeable speedup of the distributed planners; second—and, we argue, more importantly—a substantial reduction in message traffic.
| |
