Robust Imitation via Mirror Descent Inverse Reinforcement Learning
This work addresses robustness issues in imitation learning for AI agents, though it is incremental as it builds on existing adversarial frameworks.
The paper tackles the problem of uncertain reward signals in adversarial imitation learning by proposing a mirror descent-based inverse reinforcement learning method that regulates reward learning with local geometric constraints, achieving a regret bound of O(1/T) and outperforming existing methods in benchmarks.
Recently, adversarial imitation learning has shown a scalable reward acquisition method for inverse reinforcement learning (IRL) problems. However, estimated reward signals often become uncertain and fail to train a reliable statistical model since the existing methods tend to solve hard optimization problems directly. Inspired by a first-order optimization method called mirror descent, this paper proposes to predict a sequence of reward functions, which are iterative solutions for a constrained convex problem. IRL solutions derived by mirror descent are tolerant to the uncertainty incurred by target density estimation since the amount of reward learning is regulated with respect to local geometric constraints. We prove that the proposed mirror descent update rule ensures robust minimization of a Bregman divergence in terms of a rigorous regret bound of $\mathcal{O}(1/T)$ for step sizes $\{η_t\}_{t=1}^{T}$. Our IRL method was applied on top of an adversarial framework, and it outperformed existing adversarial methods in an extensive suite of benchmarks.