Adversarial Imitation Learning with Preferences

This paper extends Adversarial Imitation Learning to simultaneously utilize both demonstrations and preferences. We leverage the connection between discriminator training and density ratio estimation to incorporate preferences into the Adversarial Imitation Learning paradigm, enabling more flexible policy learning from multiple feedback types.

Recommended citation: Aleksandar Taranovic, Andras Gabor Kupcsik, Niklas Freymuth, Gerhard Neumann. (2023). "Adversarial Imitation Learning with Preferences." International Conference on Learning Representations (ICLR).
Download Paper