Decision-Theoretic and Symbolic Planning

随着人工智能规划系统被应用于更现实的问题，在替代计划中进行选择的问题出现了。规划者通常不能证明任何替代方案是否成功，因此它在选择方案时很大程度上取决于它们实现目标的可能性和效率。决策理论为做出这些选择提供了一个规范的模型，但并不像人工智能提出的那样直接适用于规划问题，更重要的是，它没有提供一个做出选择的计算框架。特别是，决策理论没有就如何生成或修改计划提供建议。因此，决策理论和符号人工智能规划算法具有互补的优势：前者提供了丰富的表示方法来表达备选方案之间的选择，并提供了如何合理地做出这些选择的理论；后者提供了一种计算方法来生成和改进计划，但要在受限的假设下进行。这两种方法将结合在一起，产生一个实用的决策理论规划者。这项研究包括理论和计算两个方面，前者旨在协调决策理论和AI问题表示，后者旨在开发一种高效地构建、评估和选择计划方案的算法。

As AI planning systems are applied to more realistic problems, the issue of choosing among alternative plans arises. The planner typically cannot prove that any alternative will or will not be successful, and so it much choose among plans on the basis of how likely they are to achieve their aims and how efficiently they do so. Decision theory supplies a normative model for making these choices, but is not directly applicable to the planning problem as AI has posed it, and more importantly does not provide a computational framework within which to make the choices. In particular, decision theory offers no advice on how to generate or modify plans. Decision theory and symbolic AI planning algorithms therefore offer complementary strengths: the former provides a rich representation for expressing choice among alternatives and a theory of how to make those choices rationally; the latter provides a computational method for generating and improving plans, but under restrictive assumptions. The two approaches will be integrated to produce a practical decision-theoretic planner. The research contains theoretical and computational; aspects, the former oriented toward reconciling the decision-theoretic and AI problem representations, the latter oriented toward developing an algorithm that efficiently builds, evaluated, and chooses among plan alternatives.