Optimizing Degree Distributions of LT-based Codes with Deep Reinforcement Learning

Abstract

Fountain codes are a modern class of error-correcting codes with a reasonably low encoding/decoding complexity and an ability that constitutes their rateless property to generate encoded symbols on demand. Fountain codes have found various applications, most notably in the areas of network communications, data storage systems, delay sensitive transmissions and continuous data delivery. Furthermore, fountain codes form the core part of alternative communications protocols, also known as a digital fountain communication protocol and its modifications, for future networks. Luby Transform (LT) codes are the first practical realization of fountain codes for binary and packet erasure channels. LT codes and the fountain codes that originate from LT codes (LT-based codes), such as Raptor or Zigzag Decodable fountain codes, utilize monomial distributions for degree selection, also known as degree distributions. Degree distributions have a significant impact on the overall efficiency of the fountain codes; therefore, we propose a novel technique that uses deep reinforcement learning to approximate/learn an optimal degree distribution (ODD). We obtain an approximation of an ODD, which we call a learned degree distribution (LDD), and evaluate the fountain codes with the LDD and other previously proposed degree distributions and observe improvements in their performance whenever the codes use the LDD.