基于机器学习的TCP拥塞控制算法识别研究  

Recognition of TCP Congestion Control Algorithms based on Machine Learning

作　　者：许乾坤 李烨[1] 董浩[1] 叶剑飞 李俊何 XU Qian-kun;LI Ye;DONG Hao;YE Jian-fei;LI Jun-he(School of Optical-Electrical and Computer Engineering,Shanghai University of Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院

出　　处：《通信技术》2019年第11期2700-2705,共6页Communications Technology

基　　金：华为技术有限公司合作项目(No.YBN2017080071)~~

摘　　要：准确识别服务器端采用的拥塞控制算法对于预防和缓解TCP网络拥塞具有重要意义,但目前已有的对拥塞控制算法的识别方法均存在一定问题。在对当前主流拥塞控制算法进行特性分析的基础上,利用接收端采集的数据提取丢包时拥塞窗口下降比例、拥塞避免阶段窗口增长函数、快速恢复阶段窗口增长函数等特征。为提高识别效率和准确率,提出一种极限学习机和随机森林相结合的算法,对服务器端的拥塞控制算法进行识别。与多种机器学习方法对比研究的结果表明,基于所构造的特征向量可实现对RENO和NEWRENO算法的识别,且所提识别新算法取得了比其它识别方法更优的识别效果。Accurate identification of the congestion control algorithm adopted by the server plays an important role in preventing and alleviating TCP network congestion.However,there still exist some problems in the existing identification methods of congestion control algorithms.Based on characteristics analysis of mainstream congestion control algorithms,several features i.e.,the multiplicative decrease parameter of congestion window,the window growth function in the congestion avoidance state,and the window growth function in the fast recovery state,are extracted by using the data collected by the receiver.In order to improve the recognition efficiency and accuracy,an algorithm combining extreme learning machine and random forest is proposed,so as to identify the congestion control algorithm used on the server.The results of comparison of with several machine learning methods and experiment indicate that based on the constructed feature vectors,the effective recognition of RENO and NEWRENO algorithms,can be realized,and that the proposed recognition algorithm can also achieve much better recognition performance than the existing recognition methods.

关 键 词：TCP拥塞控制 机器学习 算法识别 特征提取 极限学习机 随机森林 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP393[自动化与计算机技术—控制科学与工程]