Design of neural network quantizers for a distributed estimation system with communication constraints

We consider the problem of quantizer design in a distributed estimation system with communication constraints at the channels in the case where the observation statistics are unknown and one must rely on a training set. The method that we propose applies a variation of the Cyclic Generalized Lloyd Algorithm (CGLA) on every point of the training set and then uses a neural network for each quantizer to represent the training points along with their associated codewords. The codeword of every training point is initialized using a regression tree approach. Simulations show that the combined approach i.e. building the regression tree system and using its quantizers to initialize the neural networks provides an improvement over the regression tree approach except in the case of high noise variance.