Decentralized architectures for decision-making are coming to the fore as more and more distributed infrastructures are deployed to address current challenges and needs of society, including "smart grids" for energy distribution, sensor and actuator networks for ecological monitoring and control, sustainable mass transportation systems, etc. The objective is to ensure that local actions at the nodes result in coherent collective behavior of the network. Systems of this type operate in highly uncertain environments that are noisy, unpredictable and possibly subject to adversarial disturbances. The goal of this research project is to develop a comprehensive theoretical and algorithmic framework for real-time adaptive decision-making in such large-scale systems under resource and cost constraints.
Online decision-making is concerned with real-time sequential planning in the presence of model uncertainty, nonstationarity, and possibly adversarial disturbances. The investigators study a novel extension of this paradigm to decentralized settings, where the actions have to be taken at the nodes of a large network, and the nodes only have access to noisy local information. The research entails explicit consideration of a temporally varying environment with a priori unknown dynamics; analysis and application in settings with significant model uncertainty, potential unmodeled statistical dependencies in observations either across the network nodes or over time, and possible adversarial contamination of data; and accounting for the influence of the decisions and network actions on the surrounding environment. The theoretical component of the project captures the impact of decentralization on the quality of the decision-making; the algorithmic component is to develop, analyze, and implement algorithms that come as close as possible to the derived theoretical bounds.
