Learning to Draw Dynamic Agent Goals with Generative Adversarial Networks

We address the problem of designing artificial agents capable of reproducing human behavior in a competitive game involving dynamic control. Given data consisting of multiple realizations of inputs generated by pairs of interacting players, we model each agent's actions as governed by a time-varying latent goal state coupled to a control model. These goals, in turn, are described as stochastic processes evolving according to player-specific value functions depending on the current state of the game. We model these value functions using generative adversarial networks (GANs) and show that our GAN-based approach succeeds in producing sample gameplay that captures the rich dynamics of human agents. The latent goal dynamics inferred and generated by our model has applications to fields like neuroscience and animal behavior, where the underlying value functions themselves are of theoretical interest.