A Unified On-Demand Driving Behavior Model Based on Automated Machine Learning (AutoML)

An ideal driving behavior model should be able to adapt to constantly changing environments. Most of the existing models are proposed for specific traffic scenarios, so their transferability is insufficient and they may fail to face inexhaustible and unknown scenarios in the real world. This paper proposes a unified model that innovatively combines the transformer model and Automated machine learning (AutoML), which addresses the challenge of handling out-of-domain scenarios through the designed search space and flexible automatic search strategies. A naturalistic driving dataset was used to validate the model's effectiveness. The results show that the proposed model can accurately predict driving behavior in both scenarios and scenes, with the average displacement error (ADE) being controlled within 0.17 meters and the final displacement error (FDE) being within 0.35 meters. More importantly, the key parameters of the unified on-demand driving behavior model that adapt to different environments are identified. Scenarios with clear patterns, such as car-following and lane-changing scenarios in the road segment scene, expect eight heads of attention or more, while too many heads are redundant for complex scenarios taking the roundabout as an example.