Quantization for Distributed Estimation with Unknown Observation Statistics


We consider the problem of quantizer design in a distributed estimation system with communicationchannels of limited capacity in the case where the observation statistics are unknown and only a training sequence is available. We consider the scheme of Cooperative Design-Separate Encoding. We introduce an approach that is based on a generalization of regression trees. Our scheme involves growing and pruning of regression trees along with some labeling techniques for iteratively decreasing the estimation error. The labeling scheme that we introduce produces quantizers that have either connected or disconnected partition regions. Simulations show that the performance of our system is superior to that of the Lloyd-Max quantizers for each sensor and it is similar to that of Lam-Reibman quantizers that are based on known observation statistics.

Cite this paper

@inproceedings{Megalooikonomou1997QuantizationFD, title={Quantization for Distributed Estimation with Unknown Observation Statistics}, author={Vasileios Megalooikonomou}, year={1997} }