Analogical Replay for E cient Conditional Planning

Recently, several planners have been designed that can create conditionally branching plans to solve problems which involve uncertainty. These planners represent an important step in broadening the applicability of AI planning techniques, but they typically must search a larger space than non-branching planners, since they must produce valid plans for each branch considered. In the worst case this can produce an exponential increase in the complexity of planning. If conditional planners are to become usable in real-world domains, this complexity must be controlled by sharing planning e ort among branches. Analogical plan reuse should play a fundamental role in this process. We have implemented a conditional probabilistic planner that uses analogical plan replay to derive the maximum bene t from previously solved branches of the plan. This approach provides valuable guidance for when and how to merge di erent branches of the plan and exploits the high similarity between the di erent branches in a conditional plan, which have the same goal and typically a very similar state. We present experimental data in which analogical plan replay signi cantly reduces the complexity of conditional planning. Analogical replay can be applied to a variety of conditional planners, complementing the plan sharing that they may perform naturally.