重复批均值方法中的置信区间估计

在仿真实验设计以及仿真结果分析中,常常用到独立重复法和批均值法。独立重复法对多次运行的仿真结果进行特征分析,批均值法则用来对长时间运行的仿真结果进行特征分析。传统的独立重复法在数据有初始偏差时,以及批均值法在样本数据有自相关性时所给出估计的精确性会受到较大影响。重复批均值方法将独立重复法和批均值法相结合,在一定程度上克服了初始偏差和数据自相关的影响,但是运算量相应加大。考虑样本数据的初始偏差和自相关性,给出了重复批均值法的置信区间估计方法,并结合对未知稳态均值和方差估计,对3种方法进行了比较和分析。

Confidence intervals for replicated batch means estimators

In the simulation output analysis and experiment design of simulation systems, people usually use batch means (BM) and independent replications (IR). In the former case, batching observations in one long run are chosen, and in the latter, a number of smaller runs is replicated. However, for the former one the sample data are correlated, while the latter one suffers from an initialization bias at the beginning of each run. The replicated batch means (RBM), combining IR and BM, overcomes the problems of data correlation and initialization bias, but suffers from more calculation loads. Based on the existing studies, this paper gives RBM confidence interval estimation of the steady state mean considering the initial bias, and then compares these three methods above.