Predicting Driver Behavior on the Highway with Multi-Agent Adversarial Inverse Reinforcement Learning

For the implementation of autonomous or highly automated driving functions, predicting the driver behavior of the surrounding road users is highly relevant. This work investigates the prediction of driver behavior on the highway. For this purpose, different approaches, such as rule-based functions or behavior cloning models, have been proposed in the literature. In this work, a realization with Adversarial Inverse Reinforcement Learning (AIRL) is implemented. AIRL combines behavior cloning and reinforcement learning approaches. Based on Generative Adversarial Networks (GAN), it uses competition between a discriminator and a generator. This paper uses Proximal Policy Optimization (PPO) as the generator in AIRL. Further, the prediction is realized in a multi-agent fashion to achieve the the most accurate interaction-aware prediction. Our evaluation shows that this work achieves better accuracy than rule-based models and comparable accuracy to the state-of-the-art methods, but additionally has an extremely low inference time.