基于图卷积神经网络的智能路由算法

使用特定数学模型的路由转发算法难以满足用户多样化的服务质量需求，基于深度学习的智能路由方案因具有准确性、高效性、通用性等优势，成为路由决策的发展方向。然而，目前多数智能路由算法在网络拓扑动态变化时需要重新训练，造成路由更新不及时，难以应对网络拓扑动态变化。提出一种基于图卷积神经网络（GCN）的智能路由算法。线下利用提前采集的网络信息，根据路由开销标签训练GCN智能路由模型，通过该模型输出单跳路由开销。线上采集实时信息并根据模型输出的路由开销结果对网络层路由协议进行调整，计算最小路由开销的路由路径，实现自适应网络更新。算法利用GCN的图数据结构处理不规则的网络拓扑，通过图卷积算子自动提取特征解决路由网络多属性参数提取的问题，同时引入模糊C均值算法进行网络状态离散化分析，为数据集生成标签，从而有效监督GCN模型训练。实验结果表明，该算法较ECMP、DRL-TE和SmartRoute算法路由性能更好，其平均丢包率、时延和吞吐量指标均为最优，且相较于单一的流量模式具有更强的泛化能力。

Intelligent Routing Algorithm Based on the Graph Convolutional Neural Network Framework

The routing and forwarding algorithm using a specific mathematical model struggles to meet the diversified quality of service required by users. The intelligent routing scheme based on deep learning has become the development direction of routing decision-making because of its advantages of accuracy,efficiency,and universality. However,currently,most intelligent routing algorithms need to be retrained when the network topology changes dynamically,resulting in untimely route updates and difficult-to-handle dynamic changes of network topology. In this study,an intelligent routing algorithm based on the Graph Convolutional Neural Network(GCN) framework is proposed. When offline,it uses the network information collected in advance to train the GCN intelligent routing model according to the routing overhead label. Then,the single hop routing overhead is output through the model. When online,it collects real-time information,adjusts the network layer routing protocol according to the routing overhead results output by the model,calculates the network routing path with the minimum routing overhead,and realizes automatic adaptation to network updates. The algorithm uses the graph data structure of GCN to deal with irregular network topologies and automatically extracts the features through the graph convolution operator to solve the problem of multi-attribute parameter extraction of the routing network. Meanwhile,the Fuzzy C-Means(FCM)algorithm is introduced to discretize the network state and generate labels for the data set to effectively supervise the training of GCN model. Experimental results demonstrate that the algorithm attains better routing performance than those of ECMP,DRL-TE,and SmartRoute algorithms. Furthermore,its average packet loss rate,delay,and throughput are the best among the compared algorithms,and it has stronger generalization ability compared to a single traffic mode.