Using DomainConfigurable Search Control for Probabilistic Planning.pdfVIP

Using Domain-Configurable Search Control for Probabilistic Planning Ugur Kuter and Dana Nau Department of Computer Science and Institute for Systems Research University of Maryland College Park, MD 20742, USA Abstract Domain-configurable planners have been highly We describe how to improve the performance of successful. In the AI planning competitions, such MDP planning algorithms by modifying them to planners consistently worked in most domains, use the search-control mechanisms of planners solved the most problems, and solved them fastest such as TLPlan, SHOP2, and TALplanner. In our (Bacchus 2001; Fox Long 2002). They are used in experiments, modified versions of RTDP, LRTDP, a large variety of real-world applications (Wilkins and Value Iteration were exponentially faster than 1990; Currie Tate 1991; Nau et al. 2003). the original algorithms. On the largest problems Our contributions are as follows: the original algorithms could solve, the modified ones were about 10,000 times faster. On another • We describe how to modify any forward-chaining set of problems whose state spaces were more than MDP planning algorithm, by incorporating into 14,000 times larger than the original algorithms it the search-control algorithm from any forward- could solve, the modified algorithms took only chaining domain-configurable planner. about 1/3 second.