Learning to Simulate Crowds with Crowds

Controlling agent behaviors with Reinforcement Learning is of continuing interest in multiple areas. One major focus is to simulate multi-agent crowds that avoid collisions while locomoting to their goals. Although avoiding collisions is important, it is also necessary to capture realistic anticipatory navigation behaviors. We introduce a novel methodology that includes: 1) an RL method for learning an optimal navigational policy, 2) position-based constraints for correcting policy navigational decisions, and 3) a crowd-sourcing framework for selecting policy control parameters. Based on optimally selected parameters, we train a multi-agent navigation policy, which we demonstrate on crowd benchmarks. We compare our method to existing works, and demonstrate that our approach achieves superior multi-agent behaviors.