arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2102.11270cs.LGcs.ITcs.SYeess.SYmath.ITmath.OCstat.ML
收藏

Softmax策略梯度方法可能需要指数时间才能收敛

Softmax Policy Gradient Methods Can Take Exponential Time to Converge

University of Pennsylvania(宾夕法尼亚大学);Carnegie Mellon University(卡内基梅隆大学)

Gen Li, Yuting Wei, Yuejie Chi, Yuxin Chen

摘要:

softmax策略梯度(PG)方法在softmax策略参数化下执行梯度上升,可以说是现代强化学习中策略优化的事实上的实现之一。对于$γ$ -折扣无限地平线表马尔可夫决策过程(MDP),最近取得了显着的进展,建立全局收敛的softmax PG方法在寻找一个接近最优的政策。然而,以前的结果没有描绘出清晰的依赖性的收敛速度上的显着参数,如状态空间$\mathcal{S}$和有效的地平线$\frac{1}{1-γ}$,这两个可能是过大的基数。在本文中,我们提供了关于softmax PG方法的迭代复杂性的悲观信息,尽管假设可以获得精确的梯度计算。具体来说,我们证明了softmax PG方法与步长$η$可以采取\[ \frac{1}η |\mathcal{S}|^{2^{Ω\big(\frac{1}{1-γ}\big)}} ~\text{iterations} \]收敛,即使在存在一个良性的政策初始化和初始状态分布服从探索(使分布失配系数不是非常大)。这是通过在一个精心构造的MDP只包含三个动作的算法动态特性。我们的指数下限暗示了在加速PG方法中仔细调整更新规则或强制执行适当正则化的必要性。

英文摘要:

The softmax policy gradient (PG) method, which performs gradient ascent under softmax policy parameterization, is arguably one of the de facto implementations of policy optimization in modern reinforcement learning. For $γ$-discounted infinite-horizon tabular Markov decision processes (MDPs), remarkable progress has recently been achieved towards establishing global convergence of softmax PG methods in finding a near-optimal policy. However, prior results fall short of delineating clear dependencies of convergence rates on salient parameters such as the cardinality of the state space $\mathcal{S}$ and the effective horizon $\frac{1}{1-γ}$, both of which could be excessively large. In this paper, we deliver a pessimistic message regarding the iteration complexity of softmax PG methods, despite assuming access to exact gradient computation. Specifically, we demonstrate that the softmax PG method with stepsize $η$ can take \[ \frac{1}η |\mathcal{S}|^{2^{Ω\big(\frac{1}{1-γ}\big)}} ~\text{iterations} \] to converge, even in the presence of a benign policy initialization and an initial state distribution amenable to exploration (so that the distribution mismatch coefficient is not exceedingly large). This is accomplished by characterizing the algorithmic dynamics over a carefully-constructed MDP containing only three actions. Our exponential lower bound hints at the necessity of carefully adjusting update rules or enforcing proper regularization in accelerating PG methods.

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。

↑