Modeling Human Driving Behavior Through Generative Adversarial Imitation Learning

An open problem in autonomous vehicle safety validation is building reliable models of human driving behavior simulation. This work presents an approach to learn neural policies from real world demonstration data. We model as a sequential decision making that characterized by non-linearity and stochasticity, unknown underlying cost functions. Imitation learning for generating intelligent when the function or difficult specify. Building upon inverse reinforcement (IRL), Generative Adversarial Learning (GAIL) aims provide effective imitation even problems with large continuous state action spaces, such modeling driving. article describes use GAIL learning-based driver modeling. Because inherently multi-agent problem, where interaction between agents needs be modeled, this paper parameter-sharing extension called PS-GAIL tackle In addition, domain agnostic, it encode specific knowledge relevant process. Reward Augmented (RAIL), which modifies reward signal domain-specific agent. Finally, demonstrations are dependent latent factors may not captured GAIL. Burn-InfoGAIL, allows disentanglement variability demonstrations. experiments performed using NGSIM, real-world highway dataset. Experiments show these modifications can successfully behavior, accurately replicating realistic, emergent traffic flow arising agents.