基于机器学习替代模型分析的主体模型参数空间搜索及校准  

Parameter Space Exploration and Calibration of Agent-based Model by Machine Learning Surrogates

作　　者：张磊 ZHANG Lei(Harbin Institute of Finance,Department of Management,Harbin 150030,China)

机构地区：[1]哈尔滨金融学院管理系,黑龙江哈尔滨150030

出　　处：《系统工程》2021年第6期146-155,共10页Systems Engineering

基　　金：黑龙江省省属本科高校基本科研业务费科研项目(2019-KYYWF-003)。

摘　　要：基于主体的建模方法是研究复杂系统的有效方法。为了更为有效地描述真实系统的涌现行为,主体建模往往包含大量需要验证的参数,导致由于参数膨胀所引起的维数灾害问题越发突出,这给一般研究者的工作带来了困难,也导致对主体模型的适用性和解释能力提出了怀疑。为解决这一难题,提出了一种将机器学习和智能采样相结合的替代分析方法,在计算成本和数据容量有限的条件下显著地提升了参数估计效率。利用xgboost机器学习算法,可以对模型的所有参数按对结果影响的重要性进行排序并校准,进而进行主体建模的验证和输出结果分析等工作。将以上研究方法应用于金融学领域中的异质代理模型,取得了良好的拟合效果和预测精度,从而为复杂系统主体建模分析的参数验证问题提供了实用的技术分析手段。It is an effective way to analyze complex systems with agent-based modeling method. In order to describe the emergence behavior of real systems more effectively, agent-based modeling often contains a large number of parameters that need to be verified, which leads to the problem of dimension disaster caused by parameter expansion. This problem leads to the difficulties to the work of researchers, as well as doubts about the applicability and explanatory ability of agent-based model. This paper explicitly tackles parameter space exploration and calibration of agent-based models by combining machine-learning and intelligent iterative sampling. The proposed approach learns a fast surrogate meta-model using a limited number of agent-based models evaluations and approximates the relationship between agent-based models inputs and outputs. Performance is evaluated on the heterogeneous asset pricing model. By using machine learning algorithm of xgboost, the surrogate model can automatically search for the parameters that have an important impact on the output and keep approaching their true values, thus computing the calibration, validation and sensitivity analysis of agent modeling at a relative lower time cost.

关 键 词：主体建模 参数校准 机器学习 替代模型 异质代理模型 

分 类 号：F830[经济管理—金融学]