Transformer-driven multi-agent deep reinforcement learning based point cloud video transmissions

Abstract

The point cloud videos, a medium for representing natural content in AR/VR with point clouds, have attracted a wide range of attention for its characteristics and have the potential to be the next generation of video technology. Given the high data volume, the point cloud video raises the challenge of intelligent transmission and resource scheduling in multi-user scenarios under time-varying system conditions. In this paper, we propose a multi-agent deep reinforcement learning (DRL) approach to optimize the expected long-term multi-user QoE and adopt a Field of View (FoV) prediction model with Transformer for high-accuracy FoV prediction. Over the time horizon, the proposed approach learns to select the tiles of the corresponding video in accordance with a proposed well-defined QoE model capable of quantifying users' satisfaction for transmissions in an iterative way. Under various settings, extensive numerical experiments based on real throughput data traces and different computation capabilities data demonstrate that the proposed approach is effective for long-term multi-agent point cloud video transmissions.