Neural Adaptive Video Streaming via Imitation Learning and Reinforcement Learning

In this paper, we propose an adaptive bitrate streaming (ABR) algorithm based on imitation learning and reinforcement learning, named BC-PPO ABR. Aiming at the problems of existing learning-based ABR algorithms, such as long training time and slow convergence, behavior cloning (BC) is introduced to make the policy network fit the expert strategy to achieve rapid convergence. Considering the severe network fluctuations caused by mmWave 5G, Proximal Policy Optimization (PPO) algorithm is incorporated into the training process of the policy network to limit the magnitude of updates between the old and new policies, thereby obtaining a stable and effective bitrate selection strategy. The experimental results show that BC-PPO ABR achieves better QoE compared to representative benchmark tests.