大数据下仿真筛选实验及误差控制模型和应用

Simulation screening and error control for big data

针对＂大数据＋仿真模型＋后仿真分析＂这一新型研究范式,提出基于基效应估计（EE）,Bootstrap假设检验以及伪发现率误差控制（FDR）的＂后仿真分析＂筛选实验设计方法（简称BFEE）.BFEE用于识别对仿真模型所感兴趣的响应起到最重要主效应/交互效应作用的少量关键因子.相比传统EE,BFEE保证了信息识别的统计精度和效力.与现有的仿真实验的序贯分支筛选法（SB）相比,BFEE无需任何模型假设,从而更容易适应大数据背景下的复杂数据特征.蒙特卡罗仿真的三种方法对比实验表明,BFEE在不需任何假设的条件下即可获得理想的实验效率和效力,在降低计算实验成本同时,保证了识别重要效应的精度.基于实际背景仿真案例的应用显示BFEE具有良好的实际应用价值.

＂Big data ＋ simulation model ＋ post-simulation analysis＂ is viewed as a new research paradigm in the context of big data. To investigate an appropriate post-simulation analysis tool for the new paradigm,this paper proposes a new screening methodology based on using elementary effects, Bootstrap hypotheses testing, together with false discovery rate error techniques（abbreviated BFEE）. BFEE is capable of precisely identifying and mining the most important factors among a large number of factors in simulated systems. Compared to the conventional elementary-effects-based method, BFEE guarantees accuracy and efficacy of simulation experiments. Moreover, BFEE exhibits higher flexibility and is easily adapted to the big data context in comparison with sequential bifurcation, an alternative screening method in simulation.The comparative analysis of Monte Carlo simulations indicate that the proposed method achieves desired simulation efforts and statistical precision, without the need for overmuch assumptions. The corresponding results from a real-world case study shows the potential applications in practice.