On Exploiting Structures of Classical Planning Problems: Generalizing Entanglements

Much progress has been made in the research and development of automated planning algorithms in recent years. Though incremental improvements in algorithm design are still desirable, complementary approaches such as problem reformulation are important in tackling the high computational complexity of planning. While machine learning and adaptive techniques have been usefully applied to automated planning, these advances are often tied to a particular planner or class of planners that are coded to exploit that learned knowledge. A promising research direction is in exploiting knowledge engineering techniques such as reformulating the planning domain and/or the planning problem to make the problem easier to solve for general, state-of-the-art planners. Learning (outer) entanglements is one such technique, where relations between planning operators and initial or goal atoms are learned, and used to reformulate a domain by removing unneeded operator instances. Here we generalize this approach significantly to cover relations between atoms and pairs of operators themselves, and develop a technique for producing inner entanglements. We present methods for detecting inner entanglements and for using them to do problem reformulation. We provide a theoretical treatment of the area, and an empirical evaluation of the methods using standard planning benchmarks and state-of-the-art planners.