Izvorni znanstveni članak
https://doi.org/10.7307/ptt.v37i1.650
IALight: Importance-Aware Multi-Agent Reinforcement Learning for Arterial Traffic Cooperative Contro
Lu WEI
orcid.org/0000-0002-0973-4735
; Beijing Polytechnic College, School of Information Engineering
*
Xiaoyan ZHANG
; Beijing Polytechnic College, School of Information Engineering
Lijun FAN
; Beijing Polytechnic College, School of Information Engineering
Lei GAO
; North China University of Technology, School of Computer Science and Technology
Jian YANG
; North China University of Technology, School of Computer Science and Technology
* Dopisni autor.
Sažetak
Multi-intersection cooperative control for arterial or network scenarios is a crucial issue in urban traffic management. Multi-agent reinforcement learning (MARL) has been recognised as an efficient solution and shows outperformed results. However, most existing MARL-based methods treat all intersections equally, overlooking their varying importance, such as high traffic volume, connecting multiple main roads, serving as entry or exit point for highways or commercial areas, etc. Besides, learning efficiency and practicality remain challenges. To address these issues, this paper proposes a novel importance-aware MARL-based method named IALight for traffic optimisation control. First, a normalised traffic pressure is introduced to ensure our state and reward design can accurately reflect the status of intersection traffic flow. Second, a reward adjustment module is designed to modify the reward based on intersection importance. To enhance practicality and safety for real-world applications, we adopt a green duration optimisation strategy under a cyclic fixed phase sequence. Comprehensive experiments on both synthetic and real-world traffic scenarios demonstrate that the proposed IALight outperforms the traditional and deep reinforcement learning baselines by more than 20.41% and 17.88% in average vehicle travel time, respectively.
Ključne riječi
traffic signal control; intersection importance; multi-agent reinforcement learning; arterial cooperative control
Hrčak ID:
327638
URI
Datum izdavanja:
6.2.2025.
Posjeta: 358 *