中文摘要：

针对传统的TCP协议只有收到重复的确认或者超时才意识到拥塞已经发生以及不能预测拥塞等问题,将线性神经网络应用到拥塞控制中,用往返时延RTT（Round-Trip Time）与当前TCP吞吐量作为线性神经网络的输入,并将实验中得到的样本对网络进行训练,最终得到预测拥塞与非拥塞的决策面。通过仿真发现,它能有效预测拥塞的发生,降低网络拥塞崩溃发生的概率。

英文摘要：

The traditional TCP protocol only realizes that congestion has occurred when it receives severat duplicate acknowledgements or time is out,and it doesn＇t predict the congestion. To solve this problem,this article applies the linear neural network to congestion control, it uses round-trip time and the TCP throughput as the inputs of linear neural network, and trains the neural network by the specimens collected from experiments, at last decision surface of congestion and non-congestion is got. Simulation results show that it can effectively predict the occurrence of congestion, and reduce the occurrence of congestion collapse.