Adversarial Imitation Learning With General Function Approximation: Theoretical Analysis and Practical Algorithms

Summary

This paper analyzes adversarial imitation learning (AIL) with general function approximation, which prior theory had covered only in tabular and linear settings. It introduces optimization-based AIL (OPT-AIL), pairing online reward learning with optimism-regularized policy learning, in model-free and model-based variants. The authors report polynomial expert sample and interaction complexity, and state that OPT-AIL outperforms prior deep AIL methods on several challenging tasks.