Linear Erasure Codes Construction Using Reinforcement Learning

This work is devoted to the development of a software system for solving several problems of discrete optimization of the generator matrix of a binary linear block code and evaluating its efficiency. The input data for the optimization program are the dimension of the generator matrix, the type and parameters of the channel, and the evaluation functions of the generator matrix. The output is a generator matrix of a given dimension that minimizes a given function. Reinforcement learning (RL) was used to solve the optimization problem. The evaluation of the probability of successful transmission of a message (success rate, SR) through a given communication channel was considered as a criterion for the efficiency of codes. The developed RL-FEC software system allows you to simulate various data loss conditions, calculate code efficiency, and obtain binary linear block codes with a high SR. As part of several experiments conducted with various parameters and data transmission channels, binary linear block codes were obtained, that are comparable in terms of SR with the best representatives of other coding methods.