Decision making in dense traffic uncertainty scenarios is challenging for autonomous vehicles. Compare with costly manually designed driving policy, deep Reinforcement Learning (DRL) offers considerable potential by automatically learning and improving strategies. However, automated driving systems that only rely on the physical state space can hardly perform well in complex scenarios. In this paper, a framework is present to implement model-free reinforcement learning in complex uncertainty scenarios. Supervised learning is used to infer the latent states of other drivers through subtle cues and combine it with reinforcement learning. The performance of traditional reinforcement learning algorithms is enhanced by considering the surrounding vehicle types as well as the latent states of other drivers. Some challenging tasks are demonstrated, such as merging and overtaking in dense traffic in the Simulation of Urban Mobility (SUMO) simulation platform. The results show that our approach solves the tasks well and significantly better than the baseline.
Template Reinforcement Learning for Automated Driving with Scenario Switching Inference Labeling
27.10.2023
4784282 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Automated reinforcement learning scenario variation and impact penalties
Europäisches Patentamt | 2024
|Generalized Reinforcement Learning for Automated Driving under Uncertainty
TIBKAT | 2024
|