Title:Balancing Generalization and Specialization: Offline Metalearning for Bandwidth Estimation

Abstract:User experience in real-time video applications requires continuously adjusting video encoding bitrates to match available network capacity, which hinges on accurate bandwidth estimation (BWE). However, network heterogeneity prevents a one-size-fits-all solution to BWE, motivating the demand for personalized approaches. Although personalizing BWE algorithms offers benefits such as improved adaptability to individual network conditions, it faces the challenge of data drift -- where estimators degrade over time due to evolving network environments. To address this, we introduce Ivy, a novel method for BWE that leverages offline metalearning to tackle data drift and maximize end-user Quality of Experience (QoE). Our key insight is that dynamically selecting the most suitable BWE algorithm for current network conditions allows for more effective adaption to changing environments. Ivy is trained entirely offline using Implicit Q-learning, enabling it to learn from individual network conditions without a single, live videoconferencing interaction, thereby reducing deployment complexity and making Ivy more practical for real-world personalization. We implemented our method in a popular videoconferencing application and demonstrated that Ivy can enhance QoE by 5.9% to 11.2% over individual BWE algorithms and by 6.3% to 11.4% compared to existing online meta heuristics.