arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1902.04779cs.LGstat.ML
收藏

基于线性可加特征的样本最优参数Q学习

Sample-Optimal Parametric Q-Learning Using Linearly Additive Features

Princeton University(普林斯顿大学)

Lin F. Yang, Mengdi Wang

摘要:

考虑一个马尔可夫决策过程(MDP),它允许一组状态-动作特征,这些特征可以线性地表示过程的概率转移模型。我们提出了一个参数Q学习算法,找到一个近似的最优策略,使用的样本大小成比例的特征尺寸$K$和不变的状态空间的大小。为了进一步提高其样本效率,我们利用Bellman算子的单调性和内在噪声结构,提供存在的锚状态动作,暗示隐含的非负性的特征空间。我们增加了算法的方差减少,单调性保护,和置信区间的技术。证明了对于具有折扣因子$γ\in(0,1)$的任意大规模MDP,利用$\widetilde{O}(K/ε^2(1-γ)^3)$样本转移,可以从任意初始状态以高概率找到$ε$ -最优策略.匹配的信息理论下限被证明,确认所提出的方法相对于所有参数(多对数因子)的样本最优性。

英文摘要:

Consider a Markov decision process (MDP) that admits a set of state-action features, which can linearly express the process's probabilistic transition model. We propose a parametric Q-learning algorithm that finds an approximate-optimal policy using a sample size proportional to the feature dimension $K$ and invariant with respect to the size of the state space. To further improve its sample efficiency, we exploit the monotonicity property and intrinsic noise structure of the Bellman operator, provided the existence of anchor state-actions that imply implicit non-negativity in the feature space. We augment the algorithm using techniques of variance reduction, monotonicity preservation, and confidence bounds. It is proved to find a policy which is $ε$-optimal from any initial state with high probability using $\widetilde{O}(K/ε^2(1-γ)^3)$ sample transitions for arbitrarily large-scale MDP with a discount factor $γ\in(0,1)$. A matching information-theoretical lower bound is proved, confirming the sample optimality of the proposed method with respect to all parameters (up to polylog factors).

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。

↑