Loading...

Decoding Polar Codes with Deep Learning

Ashoori, Mohammad Hossein | 2022

423 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55447 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Behroozi, Hamid; Amini, Arash
  7. Abstract:
  8. Polar codes have received much attention to the extent that they are selected as a channel coding scheme in the 5G standard. The successive cancellation list (SCL) decoder suffers from high decoding Latency and limited Throughput due to its sequential decoding nature. Another polar decoding approach is the iterative belief propagation (BP) decoder which is inherently parallel and allows for better Decoding Latency and Throughput. However, its main drawback is an error-correction performance loss compared to the CRC-aided successive cancellation list (CA-SCL) decoder. From previous works, the CRC-aided belief propagation list (CA-BPL) decoder that benefits from the parallel structure of the BP decoder, a set of factor graph permutations, and Tanner graph of CRC codes, has been able to perform well in iterative decoding of CRC-aided polar (CA-Polar) codes.In this research, We propose a new scheme for outer codes that considers the behavior of the BP decoder and the weight spectrum of codes (weight distribution of codewords) as an alternative for CRC in concatenation with Polar Codes. We show for different codes that significant improvements in the BP decoding can be achieved compared to CA-Polar codes regarding decoding performance, computational complexity, and decoding latency. Moreover, it is shown that new codes have a better code weight spectrum. We show that deep learning can efficiently improve decoding performance and latency. For a 5G polar code of length 128, having 64 information bits and with 6-bit parity for outer code, It is shown that the proposed scheme is around 0.6 dB and 0.5 dB better than that of the CA-Polar code, for the BP decoder and the BPL decoder with a list of 8 permutations, respectively, at the target frame error rate of 10^{-4}. For the BP decoder with the min-sum approximation and E_b/N_0 = 3.0 dB, We show that the proposed scheme reduces the average number of addition and comparison operations by 24% and 74%, respectively, and the average decoding latency is reduced by 24%. Moreover, the minimum distance of the concatenated code is increased from 8 to 12, and there are fewer codewords with a weight of 12.
  9. Keywords:
  10. Deep Learning ; Polar Code ; Coding Theory ; Belief Propagation Algorithm

 Digital Object List

 Bookmark