Optimization of traffic signal control using non-collaborative multiagent deep reinforcement learning

Name: Paschol Supradith Na Ayudhya

keyword: Traffic control

; Machine learning

; Deep reinforcement learning

; Signalized intersection

Abstract: Traffic congestion poses a significant global challenge that adversely affects economies and the environment. Effective Traffic Signal Control (TSC) is crucial in managing traffic flow and mitigating congestion. Recent advancements in Deep Reinforcement Learning (DRL) offer promising opportunities for optimizing TSC. In traffic management, there is growing interested in exploring the effectiveness of novel Multi-Agent Deep Reinforcement Learning (MADRL) algorithms in contrast to conventional Static Traffic Control (STC) methods. Understanding the potential benefits and limitations of MADRL compared to STC can significantly contribute to developing and enhancing traffic management strategies. However, implementing collaborative MADRL in real-world scenarios entails substantial installation and maintenance costs. As a cost-effective alternative, Non-Collaborative Multi-Agent Deep Reinforcement Learning (NC-MADRL) presents an intriguing avenue for exploration. Using SUMO (Simulation of Urban Mobility) as a traffic simulation tool adds significance to the study, allowing for realistic evaluation and providing valuable insights for developing effective traffic management strategies and signalized intersection designs. To ensure the authenticity and practicality of this study, a comprehensive examination of various real-world factors is conducted. This includes analyzing different traffic patterns, intersection spacing, and traffic volumes to understand their influence on optimizing the MADRL model. The study also investigates the impact of DRL parameters and Neural Network (NN) design on the performance of MADRL. Hyperparameter tuning significantly improved the agent's performance, achieving a remarkable 49.61 percent enhancement in reward maximization. Optimizing NN size resulted in an average 19 percent improvement in maximizing cumulative negative reward. Medium-sized networks excelled in light to moderate traffic, while small-sized networks were more suitable for heavy traffic scenarios. Regarding the impact of intersection spacing on the agent's performance, the study consistently finds that longer intersection spacing leads to a 9 percent decrease in traffic delay compared to normal spacing and is significantly higher at approximately 78 percent compared to short spacing in terms of minimizing the traffic delay. The greatest decrease in cumulative delay, averaging 61.17 percent, is achieved with long intersection spacing. MADRL's ability to minimize traffic delay is slightly diminished as spacing decreases, with 59.33 percent and 39 percent decreases compared to STC in medium and short spacing scenarios. Additionally, when comparing the overall performance of MADRL to the traditional STC approach, the study finds an average delay decrease of 53.18 percent in general traffic configuration. Moreover, MADRL algorithms showcased their superiority over STC in optimizing TSC and reducing pollution. The MADRL model outperformed STC in every traffic scenario, reducing CO2 emissions by 13.7 percent. These findings provide valuable insights for developing effective traffic management strategies and signalized intersection designs, offering significant economic and environmental benefits to cities worldwide. Further research is needed to validate the proposed model in real-world traffic scenarios

Thammasat University. Thammasat University Library

Address: BANGKOK

Email: preserv@tu.ac.th

Name: Mongkut Piantanakulchai

Role: advisor

Created: 2022

Modified: 2024-01-02

Issued: 2024-01-02

วิทยานิพนธ์/Thesis

application/pdf

DOI: 10.14457/TU.the.2022.799

eng

DegreeName: Master of Science

Level: Master's Degree

Descipline: Engineering and Technology

Grantor: Thammasat University

©copyrights Thammasat University

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	12883paschol.pdf	13.8 MB	2	2025-10-01 14:17:17

ใช้เวลา

0.030215 วินาที