Improving the Performance of a Recurrent Neural Network Convolutional Decoder

The decoding of convolutional error correction codes can be described as combinatorial optimization problem. Normally the decoding process is realized using the Viterbi Decoder, but also artificial neural networks can be used. In this paper optimizations for an existing decoding method based on an unsupervised recurrent neural network (RNN) are investigated. The optimization criteria are given by the decoding performance in terms of bit error rate (BER) and the computational decoding complexity in terms of required iterations of the optimization network. To reduce the number of iterations and to improve the decoding performance, several design parameters, like shape of the activation function and level of self-feedback of the neurons are investigated. Furthermore the initialization of the network, the use of parallel decoders and different simulated annealing techniques are discussed.