Viewport-Aware Edge Caching for 360° Video Streaming: A Compact State Actor-Critic Approach

The growing popularity of immersive media applications calls for efficient 360◦ video streaming methods. However, delivering high-resolution spherical content imposes severe bandwidth demands that challenge modern networks. While viewport-adaptive tile-based streaming mitigates these demands by transmitting only the user’s current Field of View (FoV), delayed tile fetching due to inaccurate viewport prediction can degrade the user’s Quality of Experience (QoE). Edge caching offers an effective strategy to mitigate this problem, particularly when overlapping user requests are aggregated into virtual viewports. In this paper, to optimize caching performance, we propose a novel Deep Reinforcement Learning (DRL)-based mechanism for adaptive edge caching in 360◦ video streaming. The proposed approach formulates cache management as a sequential decision-making problem, enabling dynamic optimization of tile caching and replacement policies based on evolving system states. Specifically, our framework employs an Advantage Actor-Critic (A2C) architecture combined with Generalized Advantage Estimation (GAE) to improve learning stability in non-stationary caching environments. The simulation results demonstrate that our method significantly outperforms baseline algorithms in terms of PSNR, cache hit ratio, and backhaul traffic reduction. In particular, our A2C-based caching policy achieves a 10% to 20% improvement in cache hit ratio while increasing PSNR by approximately 2 dB.