AUTO: Adaptive Congestion Control Based on Multi-Objective Reinforcement Learning...
summary
AUTO formulates congestion control as a multi-objective MDP and trains a MORL agent plus a preference adaptation model to produce the Pareto frontier of throughput-vs-delay policies and pick preferences from observed state sequences. Training uses synchronous advantage actor-critic, history states, adaptive monitor intervals, smoothing of the loss, and an early-stage short-rollout trick. Evaluations on emulators and the Internet show adaptation to 2–10 Mbps link changes every 5s, delay reductions vs baselines (up to ~25–75% in reported cases), and resilience to stochastic loss—maintaining ~88% utilization at 80% loss (~7% above the next best).
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:22The Goal of Congestion Control
- 0:53How to set the preference?
- 2:13Existing Methods
- 3:12AUTO Overview
- 3:59Congestion Control Formulation
- 5:20Policy Model Training
- 7:24Preference Adaptation Model Training
- 8:40Consistent High Performance
- 9:53Adaptive to Rapid Network Changes
- 10:26Resilience to Stochastic Packet Loss
- 11:12Fairness against different schemes
- 12:21Conclusion
related talks
20mAI on the Fast Lane: Smart Traffic Systems to Unclog our Cities
17mTowards provably performant congestion control
15mMortise: Auto-tuning Congestion Control to Optimize QoE via Network-Aware Parameter
20mEnergy Optimized Green Light Assist in Varying Traffic Scenarios Using Reinforcement Learning
13mWIP: Learning Adversarial Attacks on Adaptive Traffic Signal Control Systems...
50mNavigating Chaos: The Consequences of Compromise in Mixed-Autonomy Freeways
25mPCC Vivace: Online-Learning Congestion Control
16mExpress Lane to Efficiency and Reliability: Multi-Dimensional Control in Meta’s Express..