Özet:
Due to the increasing number of users in the communication systems, efficient spectrum usage, high-speed data transfer, and better error performance became a ne cessity. Therefore, it is aimed to design error control codes that have low error rates and have a capacity close to the Shannon’s limit. As a consequence of Erdal Arıkan’s work, polar codes, a coding technique that is theoretically proven to achieve Shan non’s limit, are introduced. After polar codes are used in fifth- generation new radio (5G NR) technology, more studies are done about the decoding of polar codes and the polar code construction. The scope of the thesis is on reinforcement learning-based polar code construction. Initially, the preliminaries of polar codes and reinforcement learning are given. Then, several reinforcement learning-based polar code construction methods are introduced. It is shown that a reinforcement learning-based method found in the literature performs weakly for long block lengths due to high complexity and therefore, two new methods are introduced to reduce the complexity. First, a method that groups the channels into clusters and predetermines some channels as frozen or information is proposed. For long block lengths, it had a better performance than the one proposed in the literature, but its performance was unsatisfactory for much longer block lengths. To further reduce the complexity, neighbor dependency is introduced to the first method. It is shown that the performance of the neighbor dependent method is better than both methods and its performance is satisfactory for longer block lengths.
