Acquiring Nearly Optimal Peer Selection Strategy through Deep Q-Network

BitTorrent is a file sharing system based on the Peer-to-Peer (P2P) technology, in which a file is divided into several fragments called pieces, and peers sharing the file mutually upload a part of acquired pieces to other peers. Although there are several known strategies to determine the target of upload such as Tit-for-Tat and random, they could not properly behave from long-term perspectives such as 10 minutes, since they merely refer to a short-term history of 20 seconds. In this paper, we propose a method to acquire a nearly optimal peer selection strategy from long-term observations through Deep Q-Network which is a variant of the Q-learning enhanced by deep neural networks. The performance of the proposed method is evaluated by simulation. The result of simulations indicates that it realize a short download time compared with strategies adopted in BitTorrent.