ADR-X: ANN-Assisted Wireless Link Rate Adaptation for Compute-Constrained Embedded...
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
System achieves 10× lower packet loss (down to 0.1–0.5% versus typical 2–10%) while running ~100× faster than prior RL work, with inference taking 170–380 μs on a 400 MHz MCU. It combines an ANN that maps wireless CSI into a wire-like PSR predictor plus a contextual multi-armed bandit for MCS selection, uses online one-step updates and federated pretraining to reach 10–15 s convergence (40 s cold start), handles bursty losses, and adds conservative retransmission guardrails and CSI feature engineering.
AI-written summary · about
chapters (8)
- 0:00ADR-X introduction
- 2:33Game engines and controllers
- 5:19Packet loss impact
- 7:24Limitations of rate adaptation
- 9:51ANN-assisted rate adaptation
- 11:18Faster ANN design
- 13:24Online learning challenges
- 15:29ADR-X performance
related talks
13mOne Protocol to Rule Them All: Wireless Network-on-Chip using Deep Reinforcement Learning
18mExploring Practical Vulnerabilities of Machine Learning-based Wireless Systems
19mAI/ML Networking Challenges: The Fast and the Finicky!
14mReimagining SerDes for Scalable AI
22mCXL-ANNS: Software-Hardware Collaborative Memory Disaggregation and Computation...
01h 11mMachine Learning Acceleration in an Embedded World
13mAUTO: Adaptive Congestion Control Based on Multi-Objective Reinforcement Learning...
40mOptimize Your AI Cloud Infrastructure: A Hardware Perspective