(A) study on machine learning-aided error correction schemes

Abstract

This dissertation studies machine learning-aided error correction schemes. Firstly, we propose a novel decoding algorithm for low-density parity-check (LDPC) codes, which addresses performance degradation in a high signal-to-noise ratio (SNR) regime, called the error floor, caused by trapping sets (TSs). When the decoder fails with a TS, it is shown that there exist paths between unsatisfied check nodes (CNs) consisting of erroneous variable nodes (VNs) in the Tanner graph. By analyzing the statistical characteristics of the reliability of these erroneous VNs, we propose a machine learning-based method to detect TSs. The TS errors are resolved by correcting errors within the detected TSs and performing additional decoding trials. Furthermore, based on the interpretation of the trained model, we propose a low-complexity detection method for TSs that effectively mitigates the high computational complexity issues posed by deep neural networks. Performance evaluation demonstrated that the proposed scheme significantly improves the error floor performance of LDPC codes. Secondly, we propose a method for designing rate-compatible codes using an autoencoder. To reduce the training complexity, we concatenate an autoencoder with classical channel codes in such a way that the autoencoder reconstructs only the parity bits of the encoded codeword. In addition, the proposed model employs hierarchical autoencoders, which enables adaptive activation of the trained autoencoders depending on the rate requirements. Training for the proposed model is performed with multi-stage learning, and the trained rate-compatible codes are analyzed as a classical punctured code, which allows us to gain insights into the operation of the trained model. Lastly, we propose a design scheme for polar codes employing reinforcement learning (RL). The proposed scheme is conducted with Reed-Muller codes, i.e., polar codes optimized for the minimum distance, as a base code. Puncturing is first performed for the base code to gain a degree of freedom, called transmission hole, and the puncturing pattern is designed to minimize the loss in the minimum distance. Subsequently, extending is performed to generate additional bits transmitted through these transmission holes without rate loss. Since there exists no efficient tool to analyze the successive cancellation list decoding (SCLD) performance of polar codes, we model the problem of designing extending patterns as a Markov decision process and optimize it with RL techniques. Moreover, we introduce a novel idea to reduce the state and action space of RL, which greatly expedites the design of the extended bits. Performance evaluation shows that the polar codes designed with the proposed scheme have notably improved SCLD performance compared to existing polar code designs.