Modeling of human group coordination

We study the coordination in a group of humans by means of experiments and simulations. Experiments with human participants were implemented in a multiclient game setting, where players move on a virtual hexagonal lattice, can observe their and other players' positions on a screen, and receive a payoff for reaching specific goals on the playing field. Flocking behavior was incentivized by larger payoffs if multiple players reached the same goal field. We choose two complementary simulation methods to explain the experimental data: a minimal cognitive force approach, based on the maximization of future movement options in the agents' local environment, and multiagent reinforcement learning (RL), which learns behavioral policies to maximize reward based on past observations. Comparison between experimental and computer simulation data suggests that group coordination in humans can be achieved through nonspecific, information-based strategies. We also find that although the RL approach can capture some key aspects of the experimental results, it achieves lower performance compared to both the cognitive force simulation and the experiment, and matches the observed human behavior less closely.