Agnostic Q-learning with Function Approximation in Deterministic Systems: Tight Bounds on Approximation Error and Sample Complexity

The current paper studies the problem of agnostic -learning with function approximation in deterministic systems where the optimal -function is approximable by a function in the class with approximation error . We propose a novel recursion-based algorithm and show that if , then one can find the optimal policy using trajectories, where is the gap between the optimal -value of the best actions and that of the second-best actions and is the Eluder dimension of . Our result has two implications: 1) In conjunction with the lower bound in [Du et al., ICLR 2020], our upper bound suggests that the condition is necessary and sufficient for algorithms with polynomial sample complexity. 2) In conjunction with the lower bound in [Wen and Van Roy, NIPS 2013], our upper bound suggests that the sample complexity is tight even in the agnostic setting. Therefore, we settle the open problem on agnostic -learning proposed in [Wen and Van Roy, NIPS 2013]. We further extend our algorithm to the stochastic reward setting and obtain similar results.
View on arXiv