Towards DRL‐Based Resource Allocation for Delay‐Aware Video Delivery in Cloud‐Edge‐Device Networks

ABSTRACT

To address the conflict between stringent real‐time requirements and limited network resources in video transmission, this paper proposes a delay‐aware cloud‐edge‐devices collaborative video distribution strategy (CED‐VDS). The proposed CED‐VDS establishes a three‐tier cloud‐edge‐terminal architecture that leverages the computing and storage capabilities of the cloud, the low‐latency processing capacity of edge nodes and the real‐time playback status of terminals. A video request scheduling and multi‐link transmission model is constructed to enable dynamic caching, adaptive transcoding and cooperative delivery. To cope with the coupled decisions across caching, scheduling and resource allocation, we design an actor–critic deep reinforcement learning agent that perceives the holistic dynamic states of network transmission and terminal playback in real time. By jointly optimizing these factors, the agent maximizes long‐term user experience while minimizing end‐to‐end playback latency. Simulation results demonstrate that CED‐VDS converges approximately 4 faster than the policy gradient baseline, improves cache hit rates by over 18% and reduces average playback interruptions by 58% compared to existing strategies.