DQ-Ladder: A Deep Reinforcement Learning-based Bitrate Ladder for Adaptive Video Streaming

A research team from TU Wien, Alpen-Adria-Universität Klagenfurt, and other institutions has introduced DQ-Ladder, a novel AI-driven framework for optimizing adaptive video streaming. The system replaces the traditional, static 'bitrate ladder'—a predefined set of bitrate-resolution pairs—with a dynamic one generated by a Deep Q-Network (DQN) agent. This agent is trained using a reward function that balances video quality (predicted via VMAF and XPSNR), decoding time, and resolution smoothness. Crucially, DQ-Ladder uses machine learning models to predict these metrics, eliminating the need for exhaustive, compute-heavy encoding runs to evaluate every possible ladder configuration.

In extensive testing using the Versatile Video Coding (VVC) standard on 750 video sequences, DQ-Ladder outperformed four baseline methods. It demonstrated a BD-rate reduction of at least 10.3% for the XPSNR quality metric compared to the standard Apple HLS ladder, meaning it can deliver the same visual quality using significantly less bandwidth. Furthermore, it reduced average decoding time by 22%, easing the computational load on end-user devices. The system also proved robust, maintaining performance even when its internal prediction models were subjected to 20% artificial noise, showing lower sensitivity to errors than competing methods.