Joint Optimization of Distributed Broadcast Quantization Systems for Classification

We develop a simulated annealing technique to jointly optimize a distributed quantization structure meant to maximize the asymptotic error exponent of a downstream classifier or detector. This distributed structure sequentially processes an input vector and exploits broadcasts to improve the best possible error exponents. The annealing approach is a robust technique that avoids local maxima and is easily tailored to a broadcast quantizer's structural constraints