Precondition Control

Exploitingdomain specific information is a promising and perhaps essential means for improving the performance of AI planning systems. A number of important issues arise when investigating the use of such information. Among the most important concern the questions of how domain specific information is to be represented and how planning algorithms can be created or modified to take advantage of such information. In this paper we explore the simple mechanism of encoding such information as extra operator preconditions. We examine what is required of the operator description language to make this feasible and provide some empirical results on the effectiveness of this approach. We find that precondition control requires a very expressive operator description language, but that when it can be applied it can yield speedups that are orders of magnitude better than other methods of encoding domain specific information.