Improving TCP in Wireless Networks with an Adaptive Machine-Learnt Classifier of Packet Loss Causes

摘要

TCP understands all packet losses as buffer overflows and reacts to such congestions by reducing its rate. In hybrid wired/wireless networks where a non negligible number of packet losses are due to link errors, TCP is unable to sustain a reasonable rate. In this paper, we propose to extend TCP Newreno with a packet loss classifier built by a supervised learning algorithm called 'decision tree boosting'. The learning set of the classifier is a database of 25,000 packet loss events in a thousand of random topologies. Since a limited percentage of wrong classifications of congestions as link errors is allowed to preserve TCP-Friendliness, our protocol computes this constraint dynamically and tunes a parameter of the classifier accordingly to maximise the TCP rate. Our classifier outperforms the Veno and Westwood classifiers by achieving a higher rate in wireless networks while remaining TCP-Friendly.