Reinforcement Learning under Model Mismatch
This addresses the problem of model mismatch in reinforcement learning for practitioners needing reliable algorithms in imperfect environments, though it appears incremental as an extension of existing robust MDP frameworks.
The paper tackles reinforcement learning when the environment model is inaccurate by extending robust MDPs to model-free settings, developing robust versions of Q-learning, SARSA, and TD-learning with convergence proofs to approximately optimal policies and value functions, and scaling them to large MDPs via function approximation with convergence guarantees for linear architectures.
We study reinforcement learning under model misspecification, where we do not have access to the true environment but only to a reasonably close approximation to it. We address this problem by extending the framework of robust MDPs to the model-free Reinforcement Learning setting, where we do not have access to the model parameters, but can only sample states from it. We define robust versions of Q-learning, SARSA, and TD-learning and prove convergence to an approximately optimal robust policy and approximate value function respectively. We scale up the robust algorithms to large MDPs via function approximation and prove convergence under two different settings. We prove convergence of robust approximate policy iteration and robust approximate value iteration for linear architectures (under mild assumptions). We also define a robust loss function, the mean squared robust projected Bellman error and give stochastic gradient descent algorithms that are guaranteed to converge to a local minimum.