Investigation on Information Set Optimization of Polar Codes Using Concatenated NNDs

One of the possible methods for obtaining reduced latency is the use of neural network decoder (NND) on polar codes. However, the number of information bits K for one NND is limited, since the training complexity increases exponentially according to K. Therefore, concatenated NNDs are promising for practical information length. In the paper, we propose an iterative approach to optimize the information set for concatenated NNDs. Specifically, we first optimize the information set within same NND. We then optimize the information set between different NNDs. This process is iteratively performed until the block error rate (BLER) performance improvement is saturated. The simulation results show that the BLER performance is enhanced by using the proposed iterative approach.