Deep reinforcement learning enhanced optimization algorithm for adaptive bitrate video streaming

Abstract

Driven by the digital era, video traffic is growing rapidly, and users’ demand for high-quality video experiences is increasing. Adaptive bitrate (ABR) algorithms, as a key technology to optimize the transmission performance of video streams, play an important role in improving the efficiency of communication networks and the quality of experience (QoE). However, existing ABR algorithms rely too much on fixed control rules and simplified environment models, which make it difficult to provide optimal performance under complex and changing physical network environments (e.g., bandwidth fluctuations, delays, and network congestion). To address these challenges, this paper proposes a new ABR algorithm, the deep reinforcement learning enhanced ABR video stream optimization algorithm (PLL-ABR), which adopts proximal policy optimization as a reinforcement learning training framework and combines the dual clipping mechanism and deep neural networks (long short-term memory and local attention mechanism) to improve the training efficiency and policy parameter optimization capability. In addition, this paper also introduces a nonlinear entropy weight dynamic adjustment mechanism to balance exploration and utilization and enhance the stability of strategy optimization. By training the neural network model through reinforcement learning, PLL-ABR can dynamically select the future video block bitrate based on the physical state of the client video player and network environment information. Through comparison experiments with seven representative ABR algorithms, the method shows significant superiority under different physical network conditions and QoE factors (bitrate utilization, rebuffering penalty, and video smoothness penalty), with an average QoE improvement of 28.50%.