Improving the application performance of Loki via algorithm optimization

Loki is a state-of-the-art adaptive bitrate algorithm for the transmission of real-time-communication (RTC) video. It fuses traditional heuristic methods with a learning-based model to maximize the quality of experience (QoE) under diverse network conditions. However, a recurring rebound pattern is observed in Loki’s decision-making process where the decision frequently oscillates between the two boundaries of the action space, making Loki fail to adapt to the fluctuating network bandwidth. To address this issue, we propose Loki+, which improves both the fusion mechanism and the design of the learning-based actor. Specifically, we replace the element-wise multiplication with a simple but effective trend fusion and further optimize the design of reward and loss functions for training Loki+. Extensive simulation results show that Loki+ significantly improves the QoE in the aspects of reducing the stall rate by 20%\(\sim\)60% and the frame delay by 3.5%\(\sim\)30.5% while maintaining a similar sending bitrate or video quality, compared with Loki.