Macroscopic Characteristics of Mixed Traffic Flow with Deep Reinforcement Learning-Controlled Autonomous Vehicles

Forum topic · 小凯 · 2026-03-29

Summary

This paper by Pankaj Kumar, Pranamesh Chakraborty, and Subrahmanya Swamy Peruru (arXiv:2603.25328) addresses the challenge of controlling autonomous vehicles (AVs) in mixed traffic, balancing safety, efficiency, comfort, fuel economy, and traffic rule compliance while capturing heterogeneous driver behavior. The authors train a TD3 deep reinforcement learning algorithm on the NGSIM highway dataset, enabling realistic interaction with human-driven vehicles. Results show that transitioning from 100% human-driven traffic to fully RL-controlled traffic increases road capacity by approximately 7.52%. Compared with the Intelligent Driver Model (IDM), the RL-based AV improves average fuel efficiency by about 28.98% at higher speeds (above 50 km/h) and by 1.86% at lower speeds (below 50 km/h). The study provides quantitative evidence that deep reinforcement learning can improve macroscopic traffic flow characteristics in mixed autonomy environments.

Paper Overview

Field: Machine Learning
Authors: Pankaj Kumar, Pranamesh Chakraborty, Subrahmanya Swamy Peruru
Published: 2026-03-26
arXiv: 2603.25328

Abstract

Controlling autonomous vehicles (AVs) in mixed traffic presents significant challenges, requiring a balance between safety, efficiency, comfort, fuel economy, and traffic rule compliance, while also capturing heterogeneous driver behavior. This paper trains a TD3 deep reinforcement learning algorithm on the NGSIM highway dataset, enabling realistic interaction with human-driven vehicles.

Key Findings

Transitioning from fully human-driven traffic to fully RL-controlled traffic increases road capacity by approximately 7.52%.
Compared with the Intelligent Driver Model (IDM), the RL-based AV improves average fuel efficiency by about 28.98% at higher speeds (above 50 km/h).
At lower speeds (below 50 km/h), the fuel efficiency improvement is 1.86%.