Automatic planning, as a sub-discipline of AI, has been around for over thirty years. While formal foundations of the field have grown increasingly sophisticated, progress has been slow in terms of applications of AI planning techniques to realistic problems. An important reason for this state of affairs is the lack of adequate understanding of search-control and efficiency issues in planning. Accordingly, the PI proposes to develop systematic approaches for improving efficiency of AI planning techniques, so as to facilitate their application to realistic planning tasks. Specifically, he will (1) study the fundamental design tradeoffs in planning algorithms, and use that understanding to design more efficient planners, (2) start a multi-pronged project to integrate speedup learning techniques, including plan reuse, search control rules, derivational replay and abstraction into partial-order planning framework, and (3) develop hybrid planning architectures capable of facilitating structured interaction between planners and other types of domain-specific specialists, including humans. Through industrial and inter-disciplinary collaborations, he proposes to evaluate the resulting planning frameworks in realistic planning tasks such as manufacturing planning.
