Exploring Fundamental Utility Tradeoffs in Plan Reuse

尽管领域独立规划技术具有吸引力的通用性，但它们的低效率严重阻碍了将它们扩展到复杂的现实世界规划领域的尝试。一个非常有前途的解决方案是，通过重用以前生成的计划来解决新的计划问题，从而使计划者能够根据经验提高性能。因此，开发有效的计划重用框架是目前自动化规划、机器学习和基于案例的推理社区中非常活跃的研究领域。这样的规划系统的成功设计需要对计划重用中的存储、检索和修改之间的基本权衡有透彻的理解。在前人的研究中，提出了一种与领域无关的规划框架PRAIR，该框架可以在分层最小承诺规划的背景下灵活修改现有规划，以解决新的规划问题。PRAIR的实验已经证明了它的潜力，通过在各种领域中逐步修改现有的计划，可以带来数量级的规划性能改进。因此，PRAIR框架有望成为研究重用效用问题的理想测试平台。本研究在PRAIR修改框架的基础上，提出了一个统一的复用框架，用于研究计划复用中涉及的效用权衡，并将该框架用于研究将复用与其他加速学习技术如抽象和搜索控制规则集成的方法。所得到的重用框架的有效性将通过对经典规划领域和更复杂的现实世界应用程序（如流程规划）进行实验来证明。这项研究的结果将有助于设计具有明显有利效用权衡的计划重用框架，从而促进将领域独立规划技术扩展到更复杂的领域。因此，这项研究将对自动化规划、机器学习和基于案例的推理社区产生重大影响。//

Despite the attractive generality of domain independent planning techniques, their inefficiency has severely hampered the attempts to scale them to complex real world planning domains. One very promising solution to this involves enabling the planner to improve performance from experience by reusing previously generated plans to solve new planning problems. Developing effective plan reuse frameworks is thus currently a very active area of research in automated planning, machine learning and case based reasoning communities. Successful design of such planning systems requires a thorough understanding of the fundamental tradeoffs between storage, retrieval and modification in plan reuse. In the previous work, a domain independent planning framework called PRAIR which facilitates flexible modification of existing plans to solve new planning problems in the context of hierarchical least commitment planning has been developed. Experiments with PRAIR have demonstrated its potential to bring about order of magnitude improvements in planning performance by incrementally modifying existing plans in a variety of domains. PRAIR framework thus promises to be an ideal test bed for investigating the issues of utility of reuse. This research, a unified reuse framework is proposed based on PRAIR modification framework, an is used to study the utility tradeoffs involved in plan reuse This framework will also be used to study ways of integrating reuse with other speedup learning techniques such as abstraction and search control rules. The effectiveness of the resulting reuse framework will be demonstrated by experimenting with the classical planning domains and more complex real world applications such as process planning. Results from this research will help to design plan reuse frameworks with demonstrably favorable utility tradeoffs thereby facilitating scaling up of domain independent planning techniques to more complex domains. This research will thus have a significant impact on automated planning, machine learning, and case based reasoning communities. //