Towards Adaptive Packet Scheduler with Deep-Q Reinforcement Learning

The traditional way to design a packet scheduler is usually based on the prior knowledge of networking environments. With advanced networking technologies, we propose a Deep-Q (DQ) learning framework for packet scheduler to take advantage of more available information. The packet scheduler optimizes the application-specific quality of service (QoS) requirements, and adapt to the changing network environment. The DQ framework integrates the online Q-learning algorithm and a deep neural network, making it applicable to problems of large size. Without any prior training or network traffic models, the DQ-based scheduler progressively learns a good policy in real-time, based directly on the available observations.