A Hybrid Deep Reinforcement Learning Routing Method Under Dynamic and Complex Traffic with Software Defined Networking

Software-defined networking (SDN) routing based on reinforcement learning (RL) is a very promising research topic in recent years, achieving better solutions comparing to traditional network routing based on mathematical models. However, with continuous increase of network complexity and scale, the RL methods show a slow convergence speed and insufficient adaptability to network changes. This leads to the major drawbacks of existing RL algorithms in modern large-scale networks, especially with complexity and dynamics features. Therefore, this paper proposed a novel RL method based on pre-trained data called PRLR, a pre-trained reinforcement learning based SDN routing method, which can effectively improve the QoS of SDN routing and improve the convergence speed of reinforcement learning. The experimental results demonstrate that our proposed PRLR outperforms the benchmarking methods in terms of multiple metrics, such as network delays, bandwidth availability, goodput ratio, as well as the convergence efficiency and works well in dynamic routing topologies.