arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2102.04692cs.LG
收藏

基于自适应多步Bootstrap的表MDP的细粒度间隙相关界

Fine-Grained Gap-Dependent Bounds for Tabular MDPs via Adaptive Multi-Step Bootstrap

Tsinghua University(清华大学);Stanford University(斯坦福大学);University of Washington(华盛顿大学)

Haike Xu, Tengyu Ma, Simon S. Du

摘要:

本文提出了一种新的无模型算法--自适应多步Bootstrap(AMB),它具有更强的间隙相关后悔界。第一个创新是通过将乐观引导与自适应多步Monte Carlo推出相结合来估计最佳$Q$函数。第二个创新点是在不受任何其他行为支配的可容许行为中选择具有最大置信区间长度的行为。我们表明,当每个状态有一个独特的最佳行动,AMB实现了一个缺口依赖的遗憾界,只有规模与次优差距的倒数的总和。相比之下,Simchowitz和Jamieson(2019)表明,由于过度探索,所有置信上限(UCB)算法都会遭受额外的$Ω\left(\frac{S}{Δ_{min}}\right)$遗憾,其中$Δ_{min}$是最小次优间隙,$S$是状态数。我们进一步表明,对于一般的MDP,AMB遭受额外的$\frac{|Z_{mul}|}{Δ_{min}}$遗憾,其中$Z_{mul}$是一组状态-动作对$(s,a)$的满足$a$是一个非唯一的最佳行动$s$。我们补充我们的上限与下限显示依赖于$\frac{|Z_{mul}|}{Δ_{min}}$是不可避免的任何一致的算法。这个下限也意味着强化学习和上下文强盗之间的分离。

英文摘要:

This paper presents a new model-free algorithm for episodic finite-horizon Markov Decision Processes (MDP), Adaptive Multi-step Bootstrap (AMB), which enjoys a stronger gap-dependent regret bound. The first innovation is to estimate the optimal $Q$-function by combining an optimistic bootstrap with an adaptive multi-step Monte Carlo rollout. The second innovation is to select the action with the largest confidence interval length among admissible actions that are not dominated by any other actions. We show when each state has a unique optimal action, AMB achieves a gap-dependent regret bound that only scales with the sum of the inverse of the sub-optimality gaps. In contrast, Simchowitz and Jamieson (2019) showed all upper-confidence-bound (UCB) algorithms suffer an additional $Ω\left(\frac{S}{Δ_{min}}\right)$ regret due to over-exploration where $Δ_{min}$ is the minimum sub-optimality gap and $S$ is the number of states. We further show that for general MDPs, AMB suffers an additional $\frac{|Z_{mul}|}{Δ_{min}}$ regret, where $Z_{mul}$ is the set of state-action pairs $(s,a)$'s satisfying $a$ is a non-unique optimal action for $s$. We complement our upper bound with a lower bound showing the dependency on $\frac{|Z_{mul}|}{Δ_{min}}$ is unavoidable for any consistent algorithm. This lower bound also implies a separation between reinforcement learning and contextual bandits.

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。

↑