Optimizing RTC bandwidth estimation with machine learning

What this post added

This post details the adoption of a machine learning (ML)-based approach for optimizing bandwidth estimation (BWE) and congestion control in Meta's real-time communication (RTC) systems. It describes the transition from a tuned Google Congestion Controller (GCC) to an ML model that characterizes network types (e.g., random vs. congestion-induced packet loss) and predicts future network conditions. The post outlines the model architecture, which combines LSTM for time-series data and dense layers for non-time-series data, and details the process of generating training samples and tuning parameters offline. Experiment results show significant improvements in reliability and quality metrics, such as reduced connection drop rates and peer video freeze percentages.