Deep Learning Based Decoding for Polar Codes in Markov Gaussian Memory Impulse Noise Channels

In previous papers, decoding schemes which did not use machine learning considered additive white Gaussian noise or memoryless impulse noise. The decoding methods applying deep learning to reduce computational complexity and decoding latency didn’t consider the impulse noise. Here, we apply the Long Short-Term Memory (LSTM) neural network (NN) decoder for Polar codes under the Markov Gaussian memory impulse noise channel, and compare its bit error rate with the existing Polar code decoders like Successive Cancellation (SC), Belief Propagation (BP) and Successive Cancellation List (SCL). In the simulation results, we first find the optimal training SNR value 4.5 dB in the Markov Gaussian memory impulse noise channel for training the proposed LSTM based Polar code decoder. The optimal training SNR value is different from that 1.5 dB in the AWGN channel. The bit error rate of the propose LSTM based Polar code decoder is one third that of the previous non-deep-learning-based decoder SC/BP/SCL in Markov Gaussian memory impulse noise channels. The execution time of the proposed LSTM-based method is 5 ~ 12 times less and thus has much less decoding latency than that of SC/BP/SCL methods because the proposed LSTM-based method has inherent parallel structure and has one shot operation.