Power-Aware Traffic Engineering via Deep Reinforcement Learning

Power-aware traffic engineering via coordinated sleeping is often formulated into Integer Programming (IP) problems, which are generally NP-hard and the computation time does not scale for large networks, causing delayed control decision making in highly-dynamic network environment. Motivated by recent advances in deep reinforcement learning (RL), we consider building intelligent systems that learn to adaptively change router's power state according to traffic dynamics. The forward propagation property of neural networks can greatly speedup power on/off decision making. Specifically, we build a closed-loop training/control system and propose novel techniques to enhance its learning ability. Evaluation shows that our system can generate a comparable power saving action within 276ms, considering both energy efficiency and network traffic load balancing.