DeepGAL: Intelligent Vehicle Control for Traffic Congestion Alleviation at Intersections

Abstract

Intersections are prone to congestion in urban areas and making competent speed plans for vehicles to efficiently utilize green time resources is significant for congestion alleviation and driving experience improvement. The key is to increase the chance of going through green lights and avoid idling at red lights. Existing works for velocity planning fail to fully consider various realistic traffic conditions, such as interfering traffic, free lane switch, and group efficiency, thus compromising effectiveness and limiting feasibility. To address all the aforementioned issues, we propose a novel method DeepGAL for vehicle control by dividing vehicles into groups, assigning a leader in each group and delivering intelligent control on leaders with deep reinforcement learning. Extensive experiments are conducted based on two real-world datasets with distinct traffic flow rates in Hangzhou, China, and the results demonstrate that DeepGAL achieves outstanding improvement over various performance metrics applied in five classic car following models considering realistic traffic conditions. Moreover, DeepGAL outperforms four state-of-the-art baseline methods over various metrics under both light and heavy real-world traffic flows. The test of DeepGAL with indeterministic traffic-signal phase and timing (SPAT) information under an adaptive traffic signal control indicates its effectiveness even with partial SPAT information. In addition, considering the practical issue that only some of the vehicles can be controlled by our scheme, we conduct simulations on different penetration levels of controlled leader vehicles, which demonstrate that even with only 10% penetration, DeepGAL can notably alleviate congestion and enhance driving experience at intersections, validating its great feasibility.