Citation

Savithramma RM, Sumathi R, Sudhira HS. J. Eng. Res. 2023; 11(1): e100017.

Copyright

(Copyright © 2023, Elsevier Publishing)

DOI

10.1016/j.jer.2023.100017

PMID

unavailable

Abstract

The delay caused due to congestion at an intersection contributes to extended travel time and environmental pollution. Currently, adaptive traffic signal controllers are in practice to manage the traffic at intersections. The advancement in technology has grabbed the attention of researchers towards optimization of signal operations by applying state-of-art techniques and Reinforcement Learning (RL) is one of the prominent machine learning techniques highly explored. The primary elements of RL are the state, action, and reward, and are modelled uniquely by various researchers. However, the size of state-space and action-space is of concern as the larger space increases the time and computational complexity, and further demands for higher computational infrastructure. Thus, the minimized state-space has a greater significance in RL based traffic signal controllers. In this view the current study is carried out with two major objectives including: reducing the waiting delay and the state-space complexity. The reinforcement learning is applied to reduce the delay and the Gradient Boosting Regression Tree (GBRT) is combined to reduce the state complexity. The GBRT model is able to predict the green time for accumulated traffic volume with a R2-score of 0.92. The proposed Ensemble Controller (EC) is compared with a standard Webster Controller (WC) to demonstrate its efficacy. The average intersection delay is 134 and 107 s respectively for WC and EC. Decrease in delay implies the higher degree of match between the allocated green time and the accumulated traffic volume. The state is represented as traffic queue size in current study, and it ranges between 10 to 105 and ‐5 to 35 respectively before and after applying GBRT. The results demonstrated the significant reduction in waiting delay and state-space with the proposed control scheme.

Language: en

Keywords

Ensemble models; Intelligent signal controller; Machine learning; Reinforcement learning