Heterogeneous Interaction Modeling With Reduced Accumulated Error for Multiagent Trajectory Prediction.

Area of Science:

• Complex Systems Science
• Artificial Intelligence
• Data Science

Background:

• Dynamical complex systems with interactive heterogeneous agents are common, such as in urban traffic and social networks.
• Modeling agent interactions is crucial for understanding system dynamics and predicting behaviors like traffic participant trajectories.
• Heterogeneous interaction modeling is less explored than homogeneous systems, leading to significant error accumulation challenges.

Purpose of the Study:

• To propose a novel method, Heterogeneous Interaction Modeling with Reduced Accumulated Error (HIMRAE), for multiagent trajectory prediction.
• To address the challenges of complex heterogeneous interactions and error accumulation in multiagent systems.

Main Methods:

• Inferred dynamic interaction graphs with directed relations and effects from historical trajectories.
• Defined a heterogeneous attention mechanism (HAM) for aggregating influence from diverse neighbors.
• Introduced graph entropy and mixup training to mitigate spatial and temporal error accumulation, respectively.

Main Results:

• The proposed HIMRAE method effectively models heterogeneous interactions in multiagent systems.
• Experimental validation on three real-world datasets demonstrated the superiority of the HIMRAE approach.
• The method successfully reduced accumulated errors in trajectory prediction.

Conclusions:

• HIMRAE offers a robust solution for trajectory prediction in complex systems with heterogeneous agents.
• The integration of dynamic interaction graphs, HAM, graph entropy, and mixup training significantly improves prediction accuracy.
• This work advances the field of multiagent interaction modeling and trajectory prediction.