Distributed Learning Algorithms for Sensor Networks

Wireless sensor networks have received significant attention in the last decade owing to their widespread use not only in monitoring the physical world but also in surveillance. The energy and communication constraints of sensor nodes, coupled with distributed processing of sensed signals, lead to challenges in developing effective methods to perform desired inference tasks such as object detection or classification. Further, the lack of well-calibrated sensors is a major obstacle for the rapid deployment of sensor networks.;This dissertation develops gossip-based learning algorithms for distributed signal processing in sensor networks. In gossip-based algorithms, sensor nodes share information with local neighbors to converge upon common knowledge about the sensed environment. Gossip-based methods allow for manageable communication among energy-constrained nodes and also accommodate changing network communication topologies.;We consider three related problems and develop gossip-based processing solutions. We first consider the problem of joint signature estimation and node calibration using distributed measurements over a large-scale sensor network. We develop a new Distributed Signature Learning and Node Calibration algorithm, called D-SLANC, which estimates the signature of a commonly-sensed source signal and simultaneously estimates calibration parameters local to each sensor node. The approach we take is to model the sensor network as a connected graph and make use of the gossip-based distributed consensus to update the estimates at each iteration of the algorithm. We prove convergence of the algorithm to the centralized data pooling solution. We also compare its performance with the Cramer-Rao bound (CRB), and study the scaling performance of both the CRB and the D-SLANC algorithm.;Secondly, we develop a gossip-based algorithm for distributed ℓ 1-optimization in a large-scale sensor network setting. Specifically, we consider sensor nodes which can measure only a part of the entire measurement vector. We formulate the ℓ1-optimization problem as quadratic optimization and develop a distributed, gossip-based algorithm using the projected-gradient approach. We analyze the performance of the proposed algorithm using synthetic data and compare it with a standard ℓ1 solver.;Third, we consider the problem of distributed classifier learning in a large-scale sensor network setting. We adopt a machine learning approach to the problem and develop a distributed, gossip-based algorithm that learns the optimal (large-margin) hyperplane separating the two classes, using the projected-gradient approach. We illustrate the performance of the proposed algorithm using both synthetic and real-world datasets.