Related Links
- Registration form (opens in new tab)
- Testbed dataset (opens in new tab)
- Emulated dataset (opens in new tab)
- GitHub repository (opens in new tab)
- Custom Tensorflow model class and ONNX conversion script (opens in new tab)
- Custom Pytorch model class and ONNX conversion script (opens in new tab)
- Offline RL baseline model (opens in new tab) and inference script (opens in new tab)
- ML model used to predict objective video quality (opens in new tab)
Previous challenge
In the previous bandwidth estimation challenge (opens in new tab) that was hosted at ACM MMSys 2021, participants were provided with a “gym” simulation environment based on network simulator 3 (NS-3) and the challenge focused on learning a bandwidth estimator using online reinforcement learning (RL). Policies trained in simulation may not be optimal when deployed in the real world because of many challenges, including, sim-to-real gap, and the misalignment between rewards computed based on network statistics and actual user perceived quality of experience. This challenge fills this gap by providing signals that measure the quality of the received audio/video streams and focusing on training the bandwidth estimator using traces from audio/video calls conducted over the internet. Nevertheless, participants are encouraged to leverage the resources from the previous challenge and to test their proposed models on the simulation environment to ensure generalization.