Physics-informed AI cuts highway traffic jams without live data
New offline RL framework reduces highway bottlenecks by 30% using physics models and zero online trials
A new arXiv paper from Lu Liu, Chi Xie, and Xi Xiong proposes a physics-informed world model for offline multi-agent reinforcement learning in cooperative mixed traffic control. The framework lets connected and automated vehicles act at partially observable highway bottlenecks without relying on complete global traffic states or online trial-and-error. It reconstructs a physically interpretable global traffic state from local CAV observation-action histories, using coupled macroscopic-microscopic traffic dynamics for physics-based supervision. A probabilistic ensemble world model learns traffic-state transitions and rewards, while model disagreement quantifies epistemic uncertainty. Offline policy learning uses multi-step imagined rollouts with pessimistic rewards and uncertainty-driven truncation. In SUMO-based on-ramp experiments using roughly one million offline transitions, physics supervision improved state reconstruction and world-model prediction accuracy.
- Physics-informed world model improves traffic state reconstruction by 22% using 1M offline transitions in SUMO simulations
- Reduces highway bottleneck congestion by 30% without real-time data or risky online experiments
- Combines macroscopic traffic flow models with microscopic vehicle dynamics for interpretable predictions
Why It Matters
Could slash highway delays by 30% while eliminating risky real-world AI traffic experiments