Reinforcement-Learning-Based Overhead Reduction for Online Fountain Codes With Limited Feedback

We investigate the application of reinforcement learning (RL) on online fountain codes, and propose two schemes to reduce the full-recovery overhead with limited feedback. First, we use RL in determining the optimal degree of coded symbols for a given number of feedback, and propose the RL-based degree determination (RL-DD), with the help of theoretical analysis of the relationship between recovery rate and buffer occupancy. Then we propose online fountain codes with no build-up phase using sectioned distribution (OFCNB-SD), where the encoder sends symbols whose degrees are sampled from different sections of an overall distribution, and the decoder is improved to utilize coded symbols that are not immediately decodable. We present theoretical analysis of OFCNB-SD, and introduce RL-based sectioned distribution (RL-SD) scheme where the sectioning of the overall distribution is optimized with RL. Simulation results show that our proposed schemes could achieve lower full-recovery overhead with limited feedback compared to existing schemes.