Reinforcement Learning Approach to Improve Transmission Control Protocol

Transmission Control Protocol(TCP) plays an important role in everyday life, right from accessing ones mails to browsing the internet. With revolutionary mechanisms to ensure safe and consistent delivery of data and reducing the loss in the data transferred, TCP has indeed paved way for a paradigm shift in the way data is delivered over a network. TCP is proven to work in traditional environments involving conventional wired transmission, with well formulated packet loss restricting mechanisms implemented in the form of congestion control techniques. It is, however, found wanting in environments which involve a degree of heterogeneity (composed of wired and wireless nodes) or in purely wireless networks, involving multimedia data transmission. The performance improvement is achieved by developing a system that can classify losses as occurring due to congestion or due to the wireless nature and consequently controlling the congestion window size. This work seeks to create such a system based on reinforcement learning, where it first learns to differentiate and then predict wireless and congestion loss and consequently, predict the ideal size of congestion window thereby increasing the throughput of the system.