CAREER: Light-weight Near-optimal Stochastic Control Policies for Information Acquisition and Exploitation

穆纳加拉博士将为无线传感器网络和网格计算机等计算和通信系统中的不确定和不精确信息问题建立理论模型，并设计算法和策略以改善它们对系统性能的影响，从而导致此类系统更有效的大规模部署。他还将开发算法，用于计算不确定环境中的动态信息获取策略，从而共同优化获取更多信息的成本，以及利用信息来提高系统性能的收益。这类问题属于随机控制理论的较大范围，众所周知，计算和存储最优控制策略(或策略)需要指数时间和空间。然而，在穆纳加拉博士将研究的情景中，系统可以容忍规划决策质量中的某些次最优，但真正的要求是政策的计算和执行极其轻量级。这激发了一系列基本控制理论问题的计算效率近似最优策略的形成和设计。PI将利用他在近似和在线算法方面的专业知识来开发新的见解，以解决随机控制理论中的重要理论问题，重点是信息获取和利用问题。与人工智能和运筹学领域开发的技术相比，开发的解决方案除了高效实施外，还将有可证明的性能保证。本项目提出的新的算法原理将为重新发展近似随机控制理论奠定基础。

Dr. Munagala will develop theoretical models for the problems ofuncertainty and imprecise information in computing and communicationsystems such as wireless sensor networks and grid computers, anddesign algorithms and policies to ameliorate their impact on systemperformance, thereby leading to more effective larger-scaledeployments of such systems.He will also develop algorithms for computing dynamic informationacquisition strategies in uncertain environments, that jointlyoptimize the cost of acquiring more information, together with thegain from exploiting the information to enhance systemperformance. This class of problems falls within the larger area ofstochastic control theory.It is well known that computing and storing the optimal controlpolicies (or strategies) needs exponential time and space. However, inthe scenarios that Dr. Munagala will investigate, the system cantolerate certain suboptimality in the quality of planning decisions,but the real requirement is that the policies be extremely lightweightto compute and execute. This motivates the formulation and design ofcomputationally efficient approximately optimal policies for an entiregamut of fundamental control theoretic problems.The PI will draw upon his expertise in approximation and onlinealgorithms to develop new insights to address important theoreticalquestions in stochastic control theory, with emphasis on informationacquisition and exploitation problems. In contrast with techniquesdeveloped in the AI and operations research communities, the solutionsdeveloped will have provable performance guarantees, in addition tobeing efficient to implement. The novel algorithmic principlesemerging from this project will lay the foundations for developing afresh theory of approximate stochastic control.