CAREER: A Decision-Theoretic Approach to Intelligent Planning and Control

这是一个为期四年的持续奖励的第一年。本课题采用决策理论方法研究智能规划与控制问题，特别是不确定性和不完全信息问题。本研究的一个重点是开发改进的算法，用于形式化为马尔可夫决策过程和/或影响图的问题。采用了一种创新的方法，将计划（或策略）表示为有限自动机，并将用于寻找顺序计划的经典AI搜索和规划技术推广到寻找采用自动机形式的更复杂的条件计划。通过以因子表示的形式利用问题结构，以及分层和分布式问题分解，这将导致更有效的规划算法。本项目的一个补充重点是将决策理论技术应用于实际问题，包括规划大型海洋图像数据库中的知识发现过程，以及并行分布式内存环境中科学计算的动态负载平衡。这项研究将产生更有效的决策理论规划和控制算法，可以解决更大更复杂的问题。它也将导致如何将这些算法和模型应用于实际应用的更好的理解。

This is the first year of funding of a 4-year continuing award. This project adopts a decision theoretic approach to problems of intelligent planning and control, especially problems that are characterized by uncertainty and imperfect information. One focus of this research is the development of improved algorithms for problems that are formalized as Markov decision processes and/or influence diagrams. An innovative approach is adopted in which plans (or policies) are represented as finite automata, and classic AI search and planning techniques for finding sequential plans are generalized to find more complex conditional plans that take the form of automata. This should lead to more efficient planning algorithms by exploiting problem structure in the form of factored representations, as well as hierarchical and distributed problem decomposition. A complementary focus of this project is the application of decision-theoretic techniques to, practical problems, including planning the process of knowledge discovery in a large oceanographic image database, and dynamic load balancing of scientific computations in a parallel, distributed memory environment. This research will result in more efficient algorithms for decision-theoretic planning and control that can solve larger and more complex problems. It will also result in an improved understanding of how to apply these algorithms and models to practical applications.