Optimization of Distributed Coding for Sources with Memory and Applications in Sensor Networks

University Of California-Santa Barbara, Santa Barbara CA

Investigators

Abstract

0728986 Rose, Kenneth U of Cal Santa Barbara Optimization of Distributed Coding for Sources with Memory and Applications in Sensor Networks Distributed source coding is strongly motivated by high-density sensor networks with promising applications in numerous diverse scientific and engineering disciplines. Energy and capacity constraints led to extensive efforts to exploit inter-sensor (spatial) correlations so as to minimize resource requirements for data communications. However, practical distributed sources virtually always exhibit considerable time correlations. A major challenge emerges due to conflicts between the objectives of exploiting temporal versus spatial correlations. The degree to which distributed source coding will be practically applicable to sensor networks crucially depends on the development of effective solutions. The project focuses on joint exploitation of temporal and spatial correlations, which requires fundamental tradeoff analysis, development of new coder paradigms, and optimization tools to handle design intricacies and system complexity constraints. The research work comprises derivation of the theoretical foundation for the approaches from source coding, estimation and information theory principles, and the development of practical algorithmic tools for optimizing distributed predictive coders to overcome several obstacles and challenges: conflicts between prediction and distributed quantization, intractability of the cost function which is riddled with local minima, instability of training procedures due to feedback through the prediction loop, impacts of channel loss on performance and design, adaptation of the cost function to account (in time and space) for significant sensed events. Relevant tools for global optimization and for stable design of predictive coders, developed by the Principal Investigator's research group, will serve as initial building blocks for the approach. Extensive multidisciplinary interest in sensor networks is leveraged for access to diverse data and real-world experimental settings, for exposure of students to a broad mix of disciplines, and for broad dissemination of results.