An Effective Learning Scheme for Weighted-BP with Parallel Permutation Decoding

Weighted-BP is a method to improve the decoding performance of belief-propagation (BP) by learning its appropriate message weights using deep learning technique. For binary primitive BCH codes, it has been reported in literature that their performance can be enhanced more through parallel decoding which is achieved by exploiting the automorphism property of the codes. In this paper, an effective learning scheme for the parallel weighted-BP is proposed. While in the conventional work the weights preliminarily learned for a single weighted-BP were utilized as is even in the scenario of parallel decoding, we show that some gain can be further obtained if we take the parallel structure into account in the learning phase.