Subgoal ordering and granularity control for incremental planning

Abstract

In this paper, we study strategies in incremental planning for ordering and grouping subproblems partitioned by the subgoals of a planning problem when each sub-problem is solved by a basic planner. To generate a rich set of partial orders for ordering subproblems, we propose a new ordering algorithm based on a relaxed plan built from the initial state to the goal state. The new algorithm considers both the initial and the goal states and can effectively order subgoals in such a way that greatly reduces the number of invalidations during incremental planning. We have also considered trade-offs between the granularity of the subgoal sets and the complexity of solving the overall planning problem. We show an optimal region of grain size that minimizes the total complexity of incremental planning. We propose an efficient strategy to dynamically adjust the grain size in partitioning in order to operate in this optimal region. We further evaluate a redundant-execution scheme that uses two different subgoal orders in order to improve the quality of the plans generated without greatly sacrificing run-time efficiency. Experimental results on using three basic planners (metric-FF, YAHSP, and LPG-TD-speed) show that our strategies are general for improving the time and quality of each of these planners across various benchmarks