A Comparative Study on Lane-changing Decision Model Using Deep learning Methods

摘要

Lane-changing decision models play a significant role in traffic flow simulation. Accurate modeling of lane-changing behavior is still a challenging task. Rules-based models are widely used to describe this behavior in the literature. However, the accuracy of the models depends on the choice of the rules. Last ten years, this behavior has been models by deep learning approaches and the accuracy is improved significantly. Even though a lot of deep learning based models have been proposed, it is still unknown which kind of approaches is the most suitable one. In this paper, a comprehensive investigation of deep learning approaches on lane-changing modeling is carried out. In particular, the lane-changing process is first presented as a classification problem, and then the widely used deep learning architectures, such as Long Short Term Memory (LSTM), Staked Auto-Encode (SAE) and Gated Recurrent Units (GRU), are introduced to solve this problem. Comparison between these methods are carried out and results show that GRU and LSTM outperform SAE, LSTM has better performance with long term sequence while GRU has higher accuracy with short term sequence.