Adversarial Inverse Reinforcement Learning reevaluated for policy imitation and transferable reward recovery.
This paper introduces a novel model-enhanced adversarial inverse reinforcement learning framework that leverages model-based reward shaping to improve performance in stochastic environments, addressing limitations of existing AIL methods in uncertain environments.
This paper argues that the effectiveness of reward transfer in Adversarial Inverse Reinforcement Learning (AIRL) is primarily influenced by the choice of the Reinforcement Learning (RL) algorithm, specifically whether it's on-policy or off-policy, rather than the previously emphasized decomposability condition.