Exploiting memory in the search for high quality plans on the planning graph

The Graphplan algorithm for generating optimal make-span plans containing parallel sets of actions remains one of the most effective ways to generate such plans. However, despite numerous enhancements, the approach is currently significantly slower than state space planners employing distance-based heuristics to generate serial plans. This is a study of strategies that employ available memory to construct a search trace, which is used to transform the depth-first, iterative deepening A* (IDA*) nature of Graphplan's search into an iterative state space view. A family of methods is presented, each of which exploits a variant of the search trace to expedite search by learning from aspects of Graphplan's iterative search. Planners in the first class studied capture low-level aspects of the episodic search experience in the search trace so as to avoid much of Graphplan's redundant search effort. The second class of planners trade off this aspect in favor of a higher degree of freedom than Graphplan in traversing the space of ‘states’ generated during regression search on the graph. Distance-based heuristics are adapted for informed search trace traversal and an augmentation of these heuristics is developed targeting planning graph search. The step-optimal version of the PEGG planner from this second class clearly dominates even a highly enhanced version of Graphplan. Decidedly more impressive speedups are achieved by leveraging the heuristic in a beam search mode on the search trace, whereby virtually optimal parallel plans can be generated at speeds competitive with a state-of-the-art heuristic state space planner. Beyond search efficiency benefits afforded by the search trace, this study examines some unique capabilities it engenders. The combination of the planning graph and search trace structures proves to be an effective substrate for finding and storing an arbitrarily large number of the plans latent in a given length planning graph. In action cost augmented domains, the multi-PEGG system succeeds in efficiently streaming plans of increasing quality, as measured by the user's weighting of multiple quality metrics. Results demonstrate that multi-PEGG can efficiently stream plans of increasing quality with respect to multiple criteria, and suggest several avenues for significant advances in this regard.