Reinforcement learning for peer to peer video streaming applications

In this study, a system with reinforcement learning for push-pull mesh based video streaming applications running over p2p networks is designed. In push-pull based video streaming systems, each node in the system may receive video data from more than one parent. In the proposed system, a node which started to receive insufficient video data from any parent selects a new parent with a probabilistic method based on its previous experience. Afterwards, it learns that parent´s performance with Q-learning. In this paper, design parameters of the proposed system is given and performance results obtained from simulations are compared with random parent selection method. The observed results show that the proposed system is superior than the random parent selection method in terms of average bitrate and the number of parent exchanges.