Efficient design of quantisers for distributed estimation

Abstract

An efficient design algorithm for local quantisers at nodes that optimises the global metric (i.e. estimation error) for distributed estimation systems is proposed. Noting that the quantisation partitions constructed to minimise the estimation error are not independently encodable at each node, a novel technique of compressing the partitions into sets of codewords while ensuring a perfect independent encoding is presented, leading to the sets with reasonable cardinality, which enables transmission of multiple codewords at nodes for estimation. Experiments demonstrate a significantly improved performance over various quantiser designs while achieving about 26% reduction in the size of codeword sets compared with a recent design.