Deep Reinforcement Learning Based Overfill Rendering, Offloading and Subband Allocation for Edge-Assisted VR System

Computation offloading in edge-assisted virtual reality (VR) relies on image reprojection to mitigate motion sickness caused by long latency, but this introduces black borders due to missing pixels, which can degrade the user experience. Overfill rendering, which renders an expanded viewport beyond its native size, offers a solution to the black border problem. However, this comes at the cost of increased pixel data size, thus necessitating optimization to balance the quality of experience and service costs. Consequently, determining the optimal overfill image size becomes closely linked to offloading decisions and wireless network management. Current overfill rendering techniques rely on a fixed scaling factor, and comprehensive optimization of edge-assisted VR systems that incorporate dynamic adaptation of overfilling factors has not yet been explored in the literature. In this context, we formulate a unified problem of overfilling, offloading, and subband allocation for an edge-assisted VR system utilizing non-orthogonal multiple access. Our solution employs a deep reinforcement-learning-based approach, which consists of the overfilling decision component, employing per-user agents, and the radio resource allocation component for determining offloading decisions and subband allocation, with the aid of motion prediction. We develop four dynamic overfilling strategies: asymmetric overfilling with different overfilling factors for different sides of the VR frame, symmetric overfilling with a single factor, and velocity-adaptive overfilling with a coefficient factor. In each strategy, the overfilling factors are determined by the overfilling decision component as its action. Our numerical evaluation, conducted using real VR motion traces, demonstrates that the proposed solution, along with the developed overfilling strategies, outperforms other benchmarks, reducing black border amounts by up to 92% while consuming comparable energy to the fixed overfilling strategy.