Routing services in next generation networks not only need to provide good transmission quality in heterogeneous network environments, but also need to meet the differentiated performance requirements of different applications. For example, real-time applications require low latency performance guarantees, while low-cost applications pay more attention to strict cost guarantees. Recently, deep learning has been widely applied in the field of network. With the aid of the powerful deep neural networks, the communication network can perform the routing operation intelligently to avoid the possible failure and congestion. However, existing deep learning-based network routing algorithms cannot satisfy the specific performance requirements of users, because this kind of algorithm is an unconstrained feature learning method essentially, while the routing requirements of different applications are really a constrained problem. In order to solve the above problems, we propose a deep learning-based constrained intelligent routing method, which combines the advantages of Lagrange multiplier method for solving constrained problems and the learning ability of deep learning methods, making the routing service can not only learn complex features to adapt to network environments, but also can meet differentiated requirement of users on the performance. To the best of our knowledge, this is the first work to solve the constrained routing problem by using deep learning system. Experimental results prove the effectiveness of the proposed method and show it is a method suitable for providing high-quality routing services for the next generation network.

下一代网络中的路由服务不仅需要在异构网络环境中提供良好的传输质量，还需要满足不同应用的差异化性能需求。例如，实时应用需要低延迟的性能保证，而低成本应用则更注重严格的成本保证。近年来，深度学习在网络领域得到了广泛的应用。借助强大的深度神经网络，通信网络可以智能地进行路由操作，避免可能出现的故障和拥塞。然而，现有的基于深度学习的网络路由算法不能满足用户特定的性能需求，因为这种算法本质上是一种无约束的特征学习方法，而不同应用的路由需求确实是一个受限的问题。为了解决上述问题，我们提出了一种基于深度学习的约束智能路由方法，它结合了拉格朗日乘子法求解约束问题的优点和深度学习方法的学习能力，使得路由服务不仅可以学习复杂功能来适应网络环境，同时也能满足用户对性能的差异化要求。据我们所知，这是第一个使用深度学习系统解决约束路由问题的工作。实验结果证明了该方法的有效性，表明它是一种适合为下一代网络提供高质量路由服务的方法。