Distributed Graphplan

Abstract

Significant advances in plan synthesis under classical assumptions have occurred in the last seven years. Such efficient planners are all centralized planners. One very major development among these is the Graphplan planner. Its popularity is clear from its several efficient adaptations/extensions. Since several practical planning problems are solved in a distributed manner it is important to adapt Graphplan to distributed planning. This involves dealing with significant challenges like decomposing the goal and set of actions without losing completeness. We report two sound two-agent planners DGP (distributed Graphplan) and IG-DGP (interaction graph-based DGP). Decomposition of goal and action set in DGP is carried out manually and in IG-DGP it is carried out automatically based on a new representation called interaction graphs. Our empirical evaluation shows that both these distributed planners are faster than Graphplan. IG-DGP is orders of magnitude faster than Graphplan. IG-DGP benefits significantly from interaction graphs which allow decomposition of a problem into fully independent subproblems under certain conditions. IG-DGP is a hybrid planner in which a centralized planner processes a problem until it becomes separable into two independent subproblems that are passed to a distributed planner This paper also shows that advances in centralized planning can significantly benefit distributed planners.