Understanding reinforcement learned crowds

Simulating trajectories of virtual crowds is a commonly encountered task in Computer Graphics. Several recent works have applied Reinforcement Learning methods to animate agents, however they often make different design choices when it comes the fundamental simulation setup. Each these with reasonable justification for its use, so not obvious what their real impact, and how affect results. In this work, we analyze some arbitrary terms impact on learning performance, as well quality resulting measured energy efficiency. We perform theoretical analysis properties reward function design, empirically evaluate using certain observation action spaces variety scenarios, usage metrics. show that directly neighboring agents’ information generally outperforms more widely used raycasting. Similarly, nonholonomic controls egocentric observations tends produce efficient behaviors than holonomic absolute observations. has significant, potentially nontrivial results, researchers should be mindful about choosing reporting them work.