Decentralized policy learning with partial observation and mechanical constraints for multiperson modeling

Area of Science:

• Artificial Intelligence
• Robotics
• Computational Biology

Background:

• Extracting rules of real-world multi-agent behaviors is challenging.
• Conventional data-driven models often lack biological plausibility and interpretability due to ignoring agent limitations.

Purpose of the Study:

• To develop biologically plausible sequential generative models for multi-agent behaviors.
• To incorporate partial observation and mechanical constraints in a decentralized manner.

Main Methods:

• Formulated as a decentralized multi-agent imitation-learning problem.
• Utilized binary partial observation and decentralized policy models.
• Employed hierarchical variational recurrent neural networks with physical and biomechanical penalties.

Main Results:

• Demonstrated effectiveness on real-world basketball and soccer datasets.
• Showcased improvements in constraint violation reduction, long-term trajectory prediction, and handling partial observations.
• Validated the biological plausibility and interpretability of predicted behaviors.

Conclusions:

• The proposed method effectively models agent cognition and body dynamics.
• The approach serves as a powerful multi-agent simulator for generating realistic trajectories from real-world data.