Trajectory Prediction Algorithm for Multi-Agent Systems Based on HOFA-Informed Neural Networks

Trajectory prediction for multiple agents is an important problem in multi-agent systems (MASs), and is widely used in the field of autonomous driving and military. In this paper, a physical model of multi-agent systems based on high-order fully actuated (HOFA) system approach is constructed to establish corresponding data set, the a trajectory prediction model is trained. The idea of physical informed neural networks (PINN) is introduced to fuse the decision-making features and the physical model features while training, and an algorithm based on high-order fully actuated informed neural networks (HOFAINN) is proposed. In order to obtain the intention prediction results, the loss of both data set and HOFA model information is calculated and utilized in model training. Structures of both controller output predictor and trajectory result predictor are designed to fit the actual model of the MASs. The trajectory predictor is trained via the data set and tested on a typical scenario. The simulation results show that the proposed predictor has a better performance on trajectory prediction.