Phalguni Nanda

Phalguni Nanda, Zaiwei Chen

In this work, we show that natural policy gradient, a core algorithm in reinforcement learning, admits an exact formulation as a smoothed and averaged form of policy iteration. Specifically, we introduce doubly smoothed policy iteration (DSPI), a Bellman-operator framework in which each policy is obtained by applying a regularized greedy step to a weighted average of past $Q$-functions. DSPI includes policy iteration, dual-averaged policy iteration, natural policy gradient, and more general policy dual averaging methods as special cases. Using only monotonicity and contraction of smoothed Bellman operators, we prove distribution-free global geometric convergence of DSPI. Consequently, standard natural policy gradient and policy dual averaging achieve an iteration complexity of $\mathcal{O}((1-γ)^{-1}\log((1-γ)^{-1}ε^{-1}))$ for computing an $ε$-optimal policy, without modifying the MDP, adding regularization beyond the mirror map inherent in the update, or using adaptive, trajectory-dependent stepsizes. For the unregularized greedy case, corresponding to dual-averaged policy iteration, we also prove finite termination. The same Bellman-operator framework further extends to discounted MDPs with linear function approximation and stochastic shortest path problems.

Phalguni Nanda, Zaiwei Chen

In this work, we present the first finite-time analysis of the Q-learning algorithm under time-varying learning policies (i.e., on-policy sampling) with minimal assumptions -- specifically, assuming only the existence of a policy that induces an irreducible Markov chain over the state space. We establish a last-iterate convergence rate for $\mathbb{E}[\|Q_k - Q^*\|_\infty^2]$, implying a sample complexity of order $O(1/ε^2)$ for achieving $\mathbb{E}[\|Q_k - Q^*\|_\infty] \le ε$, matching that of off-policy Q-learning but with a worse dependence on exploration-related parameters. We also derive an explicit rate for $\mathbb{E}[\|Q^{π_k} - Q^*\|_\infty^2]$, where $π_k$ is the learning policy at iteration $k$. These results reveal that on-policy Q-learning exhibits weaker exploration than its off-policy counterpart but enjoys an exploitation advantage, as its policy converges to an optimal one rather than remaining fixed. Numerical simulations corroborate our theory. Technically, the combination of time-varying learning policies (which induce rapidly time-inhomogeneous Markovian noise) and the minimal assumption on exploration presents significant analytical challenges. To address these challenges, we employ a refined approach that leverages the Poisson equation to decompose the Markovian noise corresponding to the lazy transition matrix into a martingale-difference term and residual terms. To control the residual terms under time inhomogeneity, we perform a sensitivity analysis of the Poisson equation solution with respect to both the Q-function estimate and the learning policy. These tools may further facilitate the analysis of general reinforcement learning algorithms with rapidly time-varying learning policies -- such as single-timescale actor--critic methods and learning-in-games algorithms -- and are of independent interest.