Reinforcement Learning-Based Speed Modulation for Scalable Freeway Traffic Management in Mixed Autonomy Networks

We propose a centralized reinforcement learning (RL) framework for managing freeway traffic flow through intelligent speed modulation by connected vehicles. Unlike traditional infrastructure-based systems, our method operates without reliance on signals or fixed control logic. By segmenting road networks into super-segments and aggregating real-time traffic states, the RL agent learns scalable policies that maximize throughput, reduce delays, and mitigate congestion. The agent is trained in the high-fidelity PTV Vissim simulator using real-world network layouts derived from OpenStreetMap, which are automatically processed into Vissim-compatible formats. Experiments conducted on a representative freeway corridor in Mainz, Germany, show that our self-regulating cars (SRC) protocol consistently outperforms traditional baselines including no control, equal time, and adaptive signal strategies across key metrics. Additionally, we evaluate a transfer learning variant (SRC-TL) and demonstrate its robustness under unseen traffic scenarios. Ablation studies further confirm the system's effectiveness in mixed autonomy settings, revealing clear performance gains even when only a subset of vehicles follows the learned policy. These results highlight the viability of intelligent, infrastructure-free traffic control and its potential integration into real-time driver-assist applications.

کلیدواژه ها:

Reinforcement Learning ، Traffic Control ، Vissim Simulation ، Intelligent Transportation Systems