arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2302.07477cs.LGmath.OCstat.ML
收藏

稳定折扣马尔可夫决策过程的最优样本复杂度

Optimal Sample Complexity of Stable Discounted Markov Decision Processes

Stanford University(斯坦福大学)

Shengbo Wang, Jose Blanchet, Peter Glynn

AI总结:

本文研究折扣马尔可夫决策过程表格强化学习的最优样本复杂度,利用方差缩减Q学习框架,证明在稳定性结构下Q函数估计和策略学习的前导复杂度为$\widetilde\Theta(H(1-\gamma)^{-2}\epsilon^{-2})$,并揭示估计与控制间的分离。

摘要:

我们研究了无限时域折扣马尔可夫决策过程中表格强化学习的最优样本复杂度。已知无限制的极小极大样本复杂度为 $\widetilde\Theta((1-\gamma)^{-3}\epsilon^{-2})$,其中 $\gamma$ 表示折扣因子,$\epsilon$ 为解误差容限。然而,这种最坏情况下的速率并未考虑运营环境中常见的稳定性结构。利用方差缩减的Q学习框架,我们证明在不施加任何额外稳定性假设的情况下,上确界范数的Q函数估计和策略学习的前导样本复杂度依赖均为 $\widetilde\Theta(H(1-\gamma)^{-2}\epsilon^{-2})$,其中 $H=|v^*|_{\mathrm{span}}$ 为最优值函数的跨度。此外,假设所有策略上的均匀混合时间上界为 $t_{\mathrm{mix}}$,则最优Q函数可以在常数平移内以样本复杂度 $\widetilde\Theta\left(t_{\mathrm{mix}}^3\epsilon^{-2}\right)$ 被估计,该复杂度与 $(1-\gamma)^{-1}$ 无关。匹配的下界确立了这些前导依赖的尖锐性,并揭示了估计与控制之间的分离:尽管Q函数可以以与地平线无关的速率学习到附加常数,但策略学习通常保留其 $(1-\gamma)^{-2}$ 的依赖。

英文摘要:

We study the optimal sample complexity of tabular reinforcement learning for infinite-horizon discounted Markov decision processes. The unrestricted minimax sample complexity is known to scale as $\widetildeΘ((1-γ)^{-3}ε^{-2})$, where $γ$ denotes the discount factor and $ε$ is the solution-error tolerance. However, this worst-case rate does not account for stability structures that commonly arise in operational environments. Using the variance-reduced Q-learning framework, we show that, without imposing any additional stability assumptions, both sup-norm Q-function estimation and policy learning have leading sample-complexity dependence $\widetildeΘ(H(1-γ)^{-2}ε^{-2})$, where $H=|v^*|_{\mathrm{span}}$ is the span of the optimal value function. Moreover, assuming a uniform mixing time upper bound $t_{\mathrm{mix}}$ over all policies, the optimal Q-function can be estimated up to a constant shift with sample complexity $\widetildeΘ\left(t_{\mathrm{mix}}^3ε^{-2}\right)$, independent of $(1-γ)^{-1}$. Matching lower bounds establish the sharpness of these leading dependencies and reveal a separation between estimation and control: although the Q-function can be learned up to an additive constant at a horizon-free rate, policy learning generally retains its $(1-γ)^{-2}$ dependence.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑