Improving concurrency in temporal plans

摘要

There is an increasing interest in solving temporal planning problems. Identification and propagation of mutual exclusion relations between actions can significantly enhance the efficiency of a planner. Current definitions of mutually exclusive actions severely restrict their concurrency. In this paper, we report on thirteen groups of permanently mutually exclusive PDDL 2.1, Level 3 actions. We report on sixteen types of potentially-conflicting interactions between two actions where concurrency may be maximized by adjusting starting time of one of the two actions. We discuss several examples where actions can overlap despite conflicting preconditions and/or effects. The processes executing these actions are mostly independent. We report on a new domain-rewriting technique called “baiting” in order to improve the concurrency in temporal plans. Baiting actions lure a temporal planner into improving concurrency. The technique involves splitting user-identified operators. We report on three types of baiting (standard, double and nested) and show their suitability for various types of action interactions. Baiting requires minimal modification to the planning code. Baiting does not increase the branching in search trees. Baiting does not affect the soundness and completeness of a temporal planner. Our empirical evaluation shows that the makespans of plans generated by efficient planner Sapa with baited domain are significantly lower than makespans of plans generated without baiting.