Research On Network Congestion Control Based On Machine Learning

In the last decades,with the explosion of multimedia applications,the network quality of service is facing a set of challenges.Network congestion control which is the base of network quality of service,then become essential and urgent to be solved.Congestion control system is mainly composed of two parts:one is the end to end transport control protocol,the other is active queue management.The research of end to end congestion control has been investigated a lot,and the machine learning technologies had been introduced in 2013.However,it's nearly impossible to prove the network quality of service by only end to end congestion control.In fact,it has many defects,for instance global TCP synchronization,the fairness between data flows and the "bufferbloat" problem.Therefore this paper mainly pay attention on active queue management and the assistance of machine learning technologies for designing active queue management.The main work consists of three parts:(1)A framework for machine learning based active queue management generator.We proposed the framework to make the design of active queue management generator more convenient,flexible,and customizable.The input of the generator has two part:one is the objectives,and the other is the description of target scenarios.Then is the design of the rule which map a state space to an action.The core part is the machine learning algorithm which learn an AQM from the simulated target network scenarios guided by the objectives.It's easy to modify the design of rule,the objectives,the target network scenarios for further research based on the frame work.(2)An instance of the active queue management generator framework.In this phrase,we proposed an instance of the framework.The design of objectives,simulated target network scenarios,control rules are specific in this instance.The generated active queue management were compared with not only typical algorithms,but recently proposed algorithms.The cubic experiments prove that machine learned active queue management can perform well,and in some case even better than human-designed algorithm.(3)Active queue management generator with fairness.In this phrase,the fairness between flows is considered in the design of active queue management.The detection and measure of fairness and the management method for fairness were introduced to the framework of active queue management generator.Meanwhile,the experiments shows the generated active queue management performs well in fairness while achieving other targets.This paper proposed a framework to make it easier to design a customized active queue management generator,and also the further research.The instance in this paper prove that machine can help designing a well performed active queue management.Further more,the fairness between flows can also be proved after the fairness management was introduced to the generator.