Effective driving decision-making significantly enhances the safety of automated commercial vehicles. Different from small passenger vehicles mainly focusing on anti-collision, the inducements of collision and rollover for commercial vehicles are coupled with each other. However, these factors are not considered together which results in a limitation in the safety performance. This paper proposes a novel comprehensive driving decision-making methodology based on deep reinforcement learning (CDDM-DRL) for automated commercial vehicles in expressway scenarios. The CDDM-DRL consists of two parts. First, a feature encoding network is designed to encode hierarchical features from traffic situations and driving conditions, which can provide more useful feature information. Then an actor–critic network incorporating ensemble methods is developed to learn and provide effective driving actions, such as whether to turn and when to turn. Finally, extensive simulations in common and challenging scenarios with different traffic densities were performed. Experimental results show that our proposed method is better than some classical DRL methods in terms of time headway, backward clearance, lateral acceleration, etc. Moreover, it can prevent collision and rollover simultaneously, and realize safe driving decision-making for automated commercial vehicles.
A Comprehensive Driving Decision-Making Methodology Based on Deep Reinforcement Learning for Automated Commercial Vehicles
Int.J Automot. Technol.
International Journal of Automotive Technology ; 25 , 6 ; 1469-1483
01.12.2024
15 pages
Aufsatz (Zeitschrift)
Elektronische Ressource
Englisch
Driving Tasks Transfer in Deep Reinforcement Learning for Decision-making of Autonomous Vehicles
ArXiv | 2020
|Automatic driving decision-making method based on deep reinforcement learning
Europäisches Patentamt | 2024
|Automatic driving decision-making system based on deep reinforcement learning
Europäisches Patentamt | 2024
|Automatic driving behavior decision-making method based on deep reinforcement learning
Europäisches Patentamt | 2020
|