Data-Driven Adaptive Quantization for Distributed Inference

点击翻译按钮获取中文摘要

This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).The objective of this research is to develop an integrated framework for rate-constrained adaptive quantization techniques with application to distributed inference in wireless sensor networks. The approach allows sensor nodes to sequentially transmit their quantized data and each individual node can adaptively change its local quantizer based on prior transmissions from other nodes. Specific goals include development of linear and nonlinear adaptive quantization schemes and distributed inference methods, such as distributed estimators and detectors, distributed consensus algorithms with quantized message passing, and distributed random field estimation methods, by exploiting adaptive quantization, graphical models and distributed optimization techniques.With respect to intellectual merit, the project addresses a fundamental challenge of quantization for distributed inference in a sensor network environment, where the optimum quantizer generally cannot be implemented due to its dependence on unknown parameters associated with the random event being monitored by the sensor network. Unlike conventional methods using fixed, data-independent and often heuristically selected quantizers, this research takes a data-driven approach where, through sensor cooperation and adaptive learning, the local quantizers are sequentially updated so as to converge to an optimum solution.With respect to broader impact, the project has the potential of solving several important distributed inference problems with bandwidth and power constraints, thereby advancing the research and development of wireless sensor networks that are expected to have significant economic and social impact. The project has an integrated research and education program aimed at the training of a diverse population of students, including those from underrepresented groups.