Convolutional decoding using recurrent neural networks

We show how recurrent neutral network (RNN) convolutional decoders can be derived. As an example, we derive the RNN decoder for 1/2 rate code with constraint length 3. The derived RNN decoder is tested in Gaussian channel and the results are compared to results of optimal Viterbi decoder. Some simulation results for other constraint length codes are also given. The RNN decoder is tested also with the punctured code. It is seen that RNN decoder can achieve the performance of the Viterbi decoder. The complexity of the RNN decoder seems to increase only polynomially, while in Viterbi algorithm the increase is exponential. Also, the hardware implementation of the proposed RNN decoder is feasible