NJIT
Design and learning-based decoding of low-density-parity-check codes for multiterminal communication
Abstract
dc:description.abstract<p>The field of information theory was initiated by Claude Shannon in 1948. According to his most notable work, data can be transmitted reliably over a noisy communication channel with high probability, as long as a suitable coding scheme is employed to protect the data from channel noise. Since then, coding theorists have made numerous attempts to design low-complexity error correcting schemes which achieve the Shannon limit.</p><p>In the 1960s, Robert Gallager introduced sparse graph-based low-density-parity-check (LDPC) codes which are practically suitable for various coding applications, mainly due to their low encoding and decoding complexity compared to classical codes (e.g., Reed-Muller codes). These attractive features have allowed LDPC codes to be widely adopted for fifth generation (5G) mobile communication systems. However, the performance of LDPC codes can be further improved by eliminating detrimental objects in the code's Tanner graph. This dissertation proposes graph-theoretic construction of LDPC codes with minimal detrimental sub-structures. Novel decoding methods, based on machine learning, are also proposed to further improve the code's performance.</p><p>Firstly, a line counting (LC) technique is proposed for designing LDPC codes for point-to-point communications. LC is based on identifying a particular class of harmful structure, namely (3, 3) absorbing sets, in the LDPC code's Tanner graph. The complexity of the LC scheme is independent of the code parameters, enabling it to significantly speed up the optimization of LDPC codes, with respect to standard methods, by minimizing the number of (3, 3) absorbing sets. LC is later extended to an adaptive LC (ALC) scheme to facilitate the construction of nested LDPC codes for multi-terminal communications.</p><p>Secondly, this dissertation contributes to the development of LDPC decoding schemes with the aid of reinforcement learning (RL). Both model-based and model-free RL techniques are considered for decoding. Experimental results indicate that, for short to moderate length LDPC codes, the proposed RL-based decoders are superior to standard LDPC decoding schemes. Furthermore, the overall decoding complexity of LDPC codes is significantly reduced using the proposed RL-based decoders, when compared to conventional LDPC decoding.</p>
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy in Electrical Engineering - (Ph.D.)
- Discipline thesis:degree_discipline
- Electrical and Computer Engineering
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Habib, Salman
- Contributors dc:contributor
-
- Joerg Kliewer
- Nirwan Ansari
- Alexander Haimovich
Subjects
dc:subject × 4Identifiers
dc:identifier.*- Repository record dc:identifier
- https://digitalcommons.njit.edu/dissertations/1791