Fixed-order decoding for vector quantization over noisy channels

This paper considers the problem of vector quantization over noisy channels with memory. In previously suggested solutions, a long channel output sequence was used at the decoder to counter the effects of channel memory (sequence-based decoding). In this paper we propose a decoder that uses a fixed number of channel outputs, i.e. a fixed-order decoder. This decoder can be realized using nonlinear regression, and any low-dimensional approximation of multi-dimensional mapping can be used in implementation. In this paper, we present simulation results obtained by using a multilayer perceptron (MLP) for regression. We compare the performance of the proposed decoder with that of the sequence-based decoder for Gauss-Markov sources as well as actual image data. Our results demonstrate that fixed-order decoders can outperform sequence-based decoders at higher channel noise levels