深度学习的异构网络流量智能控制

Heterogeneous Network Traffic Intelligent Control Based on Deep Learning

目前对网络流量的智能控制大多采用深度学习技术,然而这些算法的计算复杂度随着网络规模的增大而显著增加,所以适用的网络规模较小。为了解决该问题,提出了一种基于奖励的深度学习结构,该结构分别采用深度卷积神经网络和深度信念网络预测流量负荷值矩阵和构造最终行为矩阵。因此,最终动作空间被简化为下一目的地动作矩阵,并且计算复杂度被显著降低。实验结果表明,与传统的路由方法相比,该方法在大规模网络中的丢包率和吞吐量均有较大的提高。

At present,most intelligent control of network traffic uses deep learning technology.However,the computational complexity of these algorithms increases significantly with the increase of the network scale,so the applicable network scale is smaller.In order to solve this problem,a reward-based deep learning structure is proposed here,which uses deep convolutional neural networks and deep belief networks to predict the traffic load value matrix and construct the final behavior matrix.Therefore,the final action space is simplified to the next destination action matrix,and the computational complexity is significantly reduced.The experiment results show that compared with traditional routing methods,the method has a greater improvement in packet loss rate and throughput in large-scale networks.