DRL Optimizes Deadline-Constrained Coded Caching
A new deep reinforcement learning (DRL) approach addresses deadline-constrained coded caching, where a server merges messages to serve multiple users simultaneously via coded multicasting. The problem is critical for delay-sensitive applications like video streaming, as merging decisions impact both current and future transmissions. The proposed solution formulates the delivery as a masked discrete-action queue-state control problem and trains a graph-attention policy network using proximal policy optimization. This policy network reduces the broadcast-packet expiration ratio by 40.9% (0.208 vs. 0.352) compared to baseline methods. The work is detailed in arXiv:2605.15236.
Key facts
- Coded caching allows serving multiple users with a single coded multicast message.
- Delay-sensitive applications like video streaming require online merge decisions under strict deadlines.
- Merging can be harmful for subsequent messages even if helpful for the current one.
- The problem is formulated as a masked discrete-action queue-state control problem.
- A graph-attention policy network is trained via proximal policy optimization.
- The policy network reduces broadcast-packet expiration ratio by 40.9%.
- Expiration ratio improved from 0.352 to 0.208.
- The research is published on arXiv with ID 2605.15236.
Entities
Institutions
- arXiv