EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Evaluation of a Deep Q-Network with a Quadratic Reward Function for Two-Phase Traffic Signal Control

Muhamad Achya Arifudin, Budiman Budiman, Titan Parama Yoga, Nur Alamsyah

Computer Journal · 01.02.2027 · 5 (1) · 61-75

DOI: 10.58477/cj.v5i1.532 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computer Journal, online erschienen am 01.02.2027. Verfasst von Muhamad Achya Arifudin und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Adaptive traffic signal control can respond to changing traffic demand more effectively than fixed-time operation, but deep reinforcement learning performance depends strongly on environment design, state representation, and reward formulation. This study evaluates a tuned Deep Q-Network (DQN) with a quadratic congestion penalty for two-phase traffic signal control. A dataset-driven environment paired two street segments as conflicting approaches and applied chronological training, validation, and test splits. No external microscopic traffic simulator was used; instead, the observed demand sequences drove a custom queue-based Gymnasium environment. Hyperparameters were tuned using the validation set, while final models were evaluated across three random seeds under normal, peak, incident, and imbalanced-demand scenarios. The quadratic DQN was compared with a linear-reward DQN, a fixed 50:50 controller, and a rule-based adaptive controller. Under normal traffic, the quadratic DQN achieved a mean queue of 22.16 vehicles, a 95th-percentile queue of 35.11 vehicles, throughput of 594.24 vehicles, and a service rate of 95.42%. Compared with the linear DQN, it reduced queue imbalance by 5.15% and action changes by 12.82%, although mean queue and throughput did not improve. The rule-based controller produced the best overall performance. These findings indicate that quadratic reward shaping can improve policy balance and stability, but simpler controllers may remain competitive when training data are limited. Fair baselines and multi-seed evaluation remain essential practices.

Zitieren

APA

Arifudin, M. A., Budiman, B., Yoga, T. P., & Alamsyah, N. (2027). Evaluation of a Deep Q-Network with a Quadratic Reward Function for Two-Phase Traffic Signal Control. Computer Journal, 5(1), 61-75. https://doi.org/10.58477/cj.v5i1.532

Vancouver

Arifudin MA, Budiman B, Yoga TP, Alamsyah N. Evaluation of a Deep Q-Network with a Quadratic Reward Function for Two-Phase Traffic Signal Control. Computer Journal. 2027;5(1):61-75. doi:10.58477/cj.v5i1.532

BibTeX
@article{Arifudin2027,
  author = {Arifudin, Muhamad Achya and Budiman, Budiman and Yoga, Titan Parama and Alamsyah, Nur},
  title = {Evaluation of a Deep Q-Network with a Quadratic Reward Function for Two-Phase Traffic Signal Control},
  journal = {Computer Journal},
  year = {2027},
  volume = {5},
  number = {1},
  pages = {61--75},
  doi = {10.58477/cj.v5i1.532}
}
Kostenlos starten →