LSTM-based Path Selection for Successive Cancellation List Decoding for Short Polar Codes

Abstract

Polar code is envisioned as a promising candidate for ultra-reliable low-latency communications (URLLC) in fifth-generation (5G) communication and beyond. To decode polar code, a successive cancellation list (SCL) decoder with a large list size can provide near maximum likelihood (ML) decoding performance. However, a large list size will lead to unacceptable spatial complexity, making it impractical. When the list size is small, although the complexity is low, its performance still needs to be improved. The main reason is that the sequence features implied in log-likelihood ratio (LLR) sequences are lost during calculating path metrics used for path selection. Because of the excellent sequence feature extraction ability of the long short-term memory (LSTM) network, we propose an LSTM-based path selection mechanism to replace the path metric-based path selection mechanism in SCL. In our proposed scheme, the LSTM network selects the surviving path according to the LLR sequences corresponding to the current paths. Simulation results show the effectiveness of the proposed LSTM-based path selection mechanism.