Generalizing Adaptive Video Streaming With Mixture of Experts in Heterogeneous Wireless Networks

Adaptive video streaming has become a core technology for modern video delivery, particularly in cellular networks. However, the growing dynamics of mobile environments and the diversity of user preferences present major challenges for adaptive bitrate (ABR) algorithms. Existing approaches often struggle to maintain a balance between high in-distribution performance and strong generalization across heterogeneous network conditions and personalized Quality of Experience (QoE) demands. To address these challenges, we propose NMoEABR, a unified ABR decision-making framework that integrates a nonlinear Mixtureof- Experts (NMoE) architecture with preference-aware metareinforcement learning. Specifically, we design an NMoE-based actor network that adaptively aggregates expert policies through dynamic convolution conditioned on real-time network states, thereby enhancing robustness and cross-network generalization in a zero-hot manner. Furthermore, to mitigate convergence difficulties arising from the joint optimization of expert policies and expert-weight prediction, we introduce a preference-aware meta-RL strategy that incorporates user preference embeddings and virtual preference synthesis to stabilize meta-policy updates. Comprehensive evaluations on real-world traces and wireless testbed demonstrate that NMoEABR consistently outperforms mainstream ABR benchmarks in terms of average QoE, stability, and adaptability, particularly under unseen network conditions and diverse user preference distributions.