TrajGNAS: Heterogeneous Multiagent Trajectory Prediction Based on a Graph Neural Architecture Search

Most existing deep learning-based trajectory prediction algorithms rely on manual experience to design a prediction model that matches a specific task or scenario by repeatedly and manually adjusting its structure and parameters. This approach is not only complicated and inefficient to implement but also makes it difficult to balance the achieved inference speed and prediction accuracy. To solve this problem, a heterogeneous multiagent trajectory prediction technique based on a graph neural architecture search, referred to as TrajGNAS, is innovatively proposed in this paper for the first time; this method is able to automatically perform an end-to-end graph architecture search to obtain the optimal trajectory prediction model. To improve the interpretability of the model and its ability to heterogeneously perceive different scenarios, we design a bootstrapping mechanism based on physical and risk interactions as a way to guide the architecture search process. In addition, we construct a new Neural Architecture Search loss function called SocialMI-Loss, which integrates various factors, such as accuracy, a posteriori knowledge, and model complexity, to attain a balance between accuracy and computational complexity during the learning process of the model. A series of comparative experiments conducted on two large-scale automated driving datasets (nuScenes and Argoverse) demonstrate the superiority of our method. The experimental results show that TrajGNAS performs comparably to the state-of-the-art methods, but its searched architectures still possess lightweight characteristics.