Abstract
In the context of inverse reinforcement learning (IRL) with a single expert,adversarial inverse reinforcement learning (AIRL) serves as a foundationalapproach to providing comprehensive and transferable task descriptions.However, AIRL faces practical performance challenges, primarily stemming fromthe framework's overly idealized decomposability condition, the unclear proofregarding the potential equilibrium in reward recovery, or questionablerobustness in high-dimensional environments. This paper revisits AIRL in\textbf{high-dimensional scenarios where the state space tends to infinity}.Specifically, we first establish a necessary and sufficient condition forreward transferability by examining the rank of the matrix derived fromsubtracting the identity matrix from the transition matrix. Furthermore,leveraging random matrix theory, we analyze the spectral distribution of thismatrix, demonstrating that our rank criterion holds with high probability evenwhen the transition matrices are unobservable. This suggests that thelimitations on transfer are not inherent to the AIRL framework itself, but areinstead related to the training variance of the reinforcement learningalgorithms employed within it. Based on this insight, we propose a hybridframework that integrates on-policy proximal policy optimization in the sourceenvironment with off-policy soft actor-critic in the target environment,leading to significant improvements in reward transfer effectiveness.