ON EFFICIENCIES OF STOCHASTIC OPTIMIZATION PROCEDURES UNDER IMPORTANCE SAMPLING
ON EFFICIENCIES OF STOCHASTIC OPTIMIZATION PROCEDURES UNDER IMPORTANCE SAMPLING
复制标题
DOI：
10.1109/wsc.2018.8632321
复制
发表时间：
2018-12
期刊：
影响因子：
--
通讯作者：
H. Lam;Guangxin Jiang;M. Fu
中科院分区：
文献类型：
--
作者：
H. Lam;Guangxin Jiang;M. Fu
我们研究和比较的效率随机近似（SA）和样本平均近似（SAA）的随机优化问题的决策变量内生成的概率分布或重要性抽样时，可以应用（如在模拟优化的情况下与蒙特卡洛样本）。我们解释了SA是如何在统计上比SAA更有效，在这种情况下，行为不同于传统的情况下，SAA通常被认为是统计上最优的程序。我们支持我们的主张与理论的极小极大框架和弱对偶参数。我们还展示了我们的理论研究结果与一些简单的模拟例子。
We study and compare the efficiencies of stochastic approximation (SA) and sample average approximation (SAA) for stochastic optimization problems when the decision variables are inside the generating probability distributions or when importance sampling can be applied (as in the case of simulation optimization with Monte Carlo samples). We explain how SA is statistically more efficient than SAA in such contexts, a behavior different from conventional situations where SAA is usually held as the statistically optimal procedure. We support our claim with a theoretical minimax framework and a weak duality argument. We also demonstrate our theoretical findings with some simple simulation examples.