Stochastic Control for Decentralized Systems: A Common Information Approach

点击翻译按钮获取中文摘要

Engineering systems involve a wide array of decision-making problems. Planning the production schedule for a power generator, controlling the motion of an aircraft, or selectively activating sensors to make environmental observations, all require multiple decisions to be made over time, very often in face of uncertainties about the system and the future. The success of many technological systems can be attributed to the successful solution of their underlying decision-making problems. Many of these decision-making problems were addressed under the common assumption of centrality of information and centrality of decision-making. Modern technological and economic developments, however, have led to increasingly decentralized systems. For example, the advent of smart sensors has led to sensing and monitoring systems that can make decisions in a distributed manner, the prevalence of networks has coupled individual social and economic decisions, the push for de-regulation has resulted in competitive decision-making in power systems. In all these systems, the centralized command structure is being replaced, to a varying degree, by a decentralized one. The future success of these systems will depend on our ability to understand, analyze and solve the decentralized decision-making problems these systems create. The objective of the proposed research is to develop a systematic framework for sequential decision-making problems that arise in decentralized systems operating in dynamic and uncertain environments. The proposed research will impact decentralized and networked systems in diverse application domains including infrastructure systems like power, transportation and communication networks, sensing and surveillance systems like teams of unmanned aerial vehicles or robots as well as networked control systems.This project aims to develop a systematic theory of stochastic control for decentralized and networked systems. Such systems involve a network of agents making decisions based on different information about the environment and/or about each other. In many applications, different decision makers share a common system-wide objective and should behave as cooperating members of the same team rather than competing players in a game. In such cooperative systems with a common objective, the global optimization of this objective with respect to agent decision strategies rather than game-theoretic equilibrium identification is a more appropriate solution approach. The goal of this proposal is to develop a dynamic programming like global optimization approach for sequential, decentralized decision-making problems. In particular, the proposed research will investigate decentralized control of autonomous vehicles/agents, control of decentralized finite state systems and control of information flow in sensor networks and networked control systems. The project will also explore the connections between decentralized cooperative systems and stochastic games among players with asymmetric information. The educational impact of the proposed research will include providing graduate students with a multi- disciplinary training in stochastic control, game theory, optimization and networks, development of a new graduate level course focusing on decentralized decision-making in networked systems and supporting women PhD students in our research program.