|M.Sc Student||Cohen Lior|
|Subject||Reinforcement Learning Based Flow Control for Real-Time|
Video over Cellular Channels
|Department||Department of Computer Science||Supervisors||PROF. Nahum Shimkin|
|PROF. Shie Mannor|
|Full Thesis text|
Evolving applications such as autonomous vehicle teleoperation call for high-quality and low latency video transmission. Existing technology allows high-quality video to be transmitted over several parallel cellular channels, with suitable rate control. Leveraging these systems to vehicles in motion involves further challenges due to the higher variability of the cellular channels. Major players in these technologies are the Israeli-based company LiveU and its spinoff DriveU. Here, we present a reinforcement learning based flow control algorithm that regulates the transmission rates over the cellular channels, and examine its applicability to DriveU’s video link. To train the agent, a simulator of a cellular channel is designed based on data from real cellular channels, which was provided by LiveU. Our simulation environment produces a rich and variable set of channel types for the agent during training, which is shown to improve the agent’s generalization in a simulated evaluation experiment. Here, generalization is key for overcoming the challenging sim-to-real transfer. We evaluate our algorithm in a real-world experiment and compare it to DriveU’s current system over their challenging benchmarking route. The results show that our algorithm outperforms the current controller, which was developed by a team of experts, and suggests that current reinforcement learning algorithms, trained entirely in simulation, can be used for achieving robust and satisfactory control performance in real-world problems.