Learning to adapt: Dual-router mixture-of-experts feature fusion for preference-aware adaptive video streaming

Adaptive Bitrate (ABR) algorithms play a critical role in optimizing Quality of Experience (QoE) for video streaming applications over dynamic network environments. However, most goal-driven ABR approaches rely on fixed QoE objective functions during training, which limits their adaptability when actual user preferences diverge significantly from the predefined metrics. To address this limitation, this paper introduces a novel multi-preference-aware ABR algorithm named Dual-router Mixture-of-Experts-based Multi-Preference (DMMP), which enhances generalization by adapting to diverse QoE preferences across users. Specifically, we design a dual-preference-aware gating mechanism as the policy head of a deep reinforcement learning framework, enabling dynamic capture of heterogeneous user QoE intents. Furthermore, inspired by inverse reinforcement learning, we train a preference prediction model that perceives changing network conditions in real time, allowing for adaptive optimization target adjustment and improved generalization under varying network states and user demands. Experimental results demonstrate that DMMP outperforms several representative ABR algorithms in terms of QoE optimization, exhibiting robust generalization not only under trained preference settings but also in previously unseen preference scenarios.