In recent years, machine learning (ML) has been widely used to address software-defined networking routing problems. However, existing ML-based methods still face challenges such as the lack of representation for network links, inadequate consideration of quality-of-service metrics in routing policies and difficulties in achieving globally optimal under distributed routing control. To overcome these limitations, this paper proposes a Graph Reinforcement Learning Routing scheme (GRL-R) that simultaneously embeds nodes and links features to enable network local awareness. And a comprehensive reward function is designed with taking account of packet loss, delay and throughput, then globally optimal routing policies are offered by the combination of a centrally deployed agent and a safe learning mechanism. Finally, experiments conducted on Abilene, GEANT2 and OS3E show that the proposed method outperforms the baselines.

In the most of previous studies, social interactions between vehicles are not considered explicitly when designing decision making and motion planning module. Nevertheless, the strong interaction exists in the most of driving scenarios, especially in highway junction. This paper presents a game theoretic merging behavior control system for autonomous vehicle at on-ramp junction considering the interaction between the merging vehicle and following vehicle in the main lane. In this work, a driving style estimation is proposed to deal with the heterogeneity of driving style. Then we adopt a model predictive control (MPC) method to plan the optimal merging trajectory based on the game theoretic decision making result. Finally, simulation and Human-in-the-loop (HIL) experiment result shows the effectiveness of our approach in on-ramp merging scenario.