SGER: Decision-making In Complex Systems

PI正在从理论和实践的角度探索决策理论在大型复杂系统中的应用。从理论的角度来看，他计划使用他之前引入的合理性度量概念作为探索定性决策的工具。可信性度量概括了概率度量，并提供了一个优雅的框架，用于理解不确定性度量的哪些属性对于将该不确定性度量用于各种目的（例如，作为信念修正模型或应用贝叶斯网络技术）是必要的。PI希望，在没有完整概率分布和只有粗略效用的应用程序中，合理性度量将使他能够“微调”决策方法。就实际应用而言，PI计划扩展他在将决策理论应用于数据库查询优化方面的初步工作。当用户提出查询时，通常有许多不同的计划可用于计算答案。虽然所有的计划都能正确地计算出答案，但它们在运行时间上可能差别很大。最佳计划通常取决于某些随机变量的值（运行查询时系统有多少可用内存以及各种谓词的选择性）。当前的查询优化算法只是对这些变量使用一个特定的值（例如，期望值）。PI之前已经展示了如何修改这些算法，以允许与每个变量相关联的概率分布，以便以最少的预期运行时间计算计划。从理论上讲，这种方法应该大大优于竞争对手，但理论结果是否在实践中站得住还有待实验确定。

The PI is exploring applications of decision theory to large, complex systems, both from the theoretical and practical point of view. From the theoretical end, he plans to use his previously-introduced concept of plausibility measures as a tool for exploring qualitative decision making. Plausibility measures generalize probability measures, and provide an elegant framework for understanding what properties of an uncertainty measure are necessary to use that uncertainty measure for various purposes (e.g., as a model of belief revision or to apply the techniques of Bayesian networks). The PI hopes plausibility measures will enable him to "fine-tune" approaches to decision making in applications where one does not have complete probability distributions and only rough utilities. As far as practical applications go, the PI plans to extend his initial work on applying decision theory to query optimization in databases. When a user poses a query, there are in general many different plans that can be used to compute the answer. While all the plans will compute the answer correctly, they may differ wildly in running time. What will be the best plan will in general depend on the values of certain random variables (how much memory the system has available when the query is run and the selectivity of various predicates). Current query optimization algorithms just use a particular value (e.g., the expected value) for these variables. The PI has previously shown how to modify these algorithms to allow for there being a probability distribution associated with each of these variables in order to compute the plan with the least expected running time. In theory, this approach should substantially outperform the competition, but it remains to be determined experimentally whether the theoretical results hold up in practice.