Spectral Representation Enhanced Deep Reinforcement Learning for Adaptive Video Streaming Over Wireless Networks

With the rapid proliferation of video streaming services over wireless networks, the design of adaptive bitrate (ABR) video streaming approaches has attracted significant research attention. Recent ABR approaches typically utilize the learned characteristics of input signals to determine the transmission bitrate for the next video chunk. However, they seldom investigate the correlations among input signals in their spectral domain. Therefore we propose a deep reinforcement learning based ABR approach that leverages both the input signals and their spectral representations to maximize user quality of experience. Finally, experimental results demonstrate that our method achieves an average QoE improvement of 2%–18% over previously ABR approaches.