Quantization for multiterminal inference

Abstract

This thesis deals with the topic of quantization for multiterminal detection or distributed hypothesis tests. We extend the concept of Neyman-Pearson detection and Bayesian detection to a multiterminal setting. Specifically, the scenario consists of a number of sensors which collect and process local measurements. Each sensor calculates a local statistic of the measurements, e.g., a log-likelihood ratio (LLR). The LLR value is quantized and afterwards reported to a central processing unit, the fusion center (FC). The FC gathers all LLRs from the sensors and makes a final decision. Special emphasizes is put on the issue of quantization. The local LLR values have to be quantized before they can be transmitted to the FC. Because it is not evident how to design the quantizer to maximize the overall system performance, we consider three different quantizer design strategies: The Lloyd-Max quantizer, the maximum output entropy (MOE) quantizer, and a quantizer based on the information bottleneck method (IBM). In the last part of this thesis we corroborate our theoretical findings by numerical simulation results. More precisely, we evaluate the performance of distributed hypothesis tests for a variety of different settings. Furthermore, we discuss several possible extensions which could yield improved performance.