BANQ: BayesOpt-Based Automatic Non-Uniform Quantization for SCL Polar Decoding

Abstract

Successive cancellation list (SCL) polar decoding is well-known for its outstanding error correction performance. To achieve a better balance between performance and complexity in SCL decoders, non-uniform quantization (NUQ) is commonly employed. NUQ strategically adjusts the quantization steps to improve the decoder’s performance within specific bitwidth constraints. However, existing NUQ schemes rely on the designers’ expertise and lack an automated design methodology. To address this issue, this letter formulates the NUQ optimization problem and solves it using Bayesian-optimization (BayesOpt), resulting in an automated NUQ scheme termed BANQ. Utilizing BayesOpt to fine-tune the quantization steps, BANQ empowers the SCL decoder to deliver enhanced error correction performance with comparable complexity. For a $(512,410)~5$ G polar code, compared to uniform quantization (UQ), BANQ cuts bitwidth by 20% and bit operations (BOPs) by 19.5%, achieving near floating-point accuracy. Furthermore, a case study verifies that BANQ can be effectively adapted for 5G LDPC min-sum decoding.