ITR: Distributed Learning in Sensor Networks

Princeton University, Princeton NJ

Investigators

Abstract

The possibility of deploying a large number of networked sensors presents great opportunities for a host of commercial, military, and homeland security applications, but also presents enormous technical challenges. To fully utilize the potential of such networks, advances will be required on a number of fronts and through all layers of the network. However, in addition to dealing with most of the difficult issues of wireless communication networks, sensor networks give rise to a number of additional issues as well. Sensor networks must do a great deal more than just support communication. Transporting bits does not automatically lead to intelligent decision-making, and connectivity does not automatically result in coordination. In a sensor network there is a joint purpose to be accomplished by the entire network as a whole. There are fundamental questions about which bits to transmit, where to send them, and how to utilize them. Moreover, there are issues of whether to do computations locally, or to pass information to higher layers and perform partially centralized computations. These tasks must be accomplished in the face of scarce resources, notably limited time, bandwidth, and power. One key area where breakthroughs are needed, and which is the focus of this project, concerns how to learn, adapt, and make decisions to carry out the goals of the sensor network in a complex and distributed environment. Accomplishing a joint goal for the entire network poses significant information processing challenges. Myopic or local information gathered by the individual sensors must be fused for global decision-making. These tasks are complicated by the large number of sensors, each of which may be heterogeneous, multimodal, possibly dynamic, and uncalibrated. In some applications the scene-sensor geometry may be unknown or only partially known, and the environment itself will typically be complex and dynamic. The key element that distinguishes learning, adaptation, and decision-making in sensor networks with most previous work in these areas is the distributed and local nature of the information gathering together with the rich and varied structure in the information and decisions to be made in the face of limited resources. A degenerate appeal to existing methods by sending all the data to a centralized node may be impossible or infeasible in many situations, and certainly will not be the most effective way to utilize limited resources. In this project, we address key problems in distributed learning in sensor networks, to begin to develop an understanding of the fundamental capabilities and limitations of learning and decision-making in sensor networks.