Using fuzzy inference to improve TCP congestion control over wireless networks

While modern wireless networks have been in development for a couple of decades, the Transmission Control Protocol (TCP) which runs over those networks has existed since the mid 1970s. As it was developed before wireless networks were even conceived, TCP was not optimized to consider the physical characteristics of wireless links. Specifically, TCP responds to packet loss due to link errors in the same way it responds to packet loss due to congestion: it cuts back the rate at which traffic is sent. A means of improving performance of TCP over wireless links is to classify packet losses, and react only to those losses perceived as being caused by network congestion. There is a demonstrated applicability of fuzzy inference in solving problems that are difficult to stochastically model or analyze. Fuzzy inference systems allow problems to be defined intuitively using a propositional IF THEN rule base. In this paper, we use environmental variables available to TCP implementations to feed a fuzzy inference system that classifies packet loss due to congestion or wireless problem without sacrificing the end-to-end reliability of TCP. Using the network simulator ns-3 we demonstrate that our TCP+FUZZY Classifier implementation performs better than de facto TCP implementations on the Internet while maintaining TCP-friendliness property.