基于深度学习的智能全局灵敏度分析

Intelligent Global Sensitivity Analysis Based on Deep Learning

提出一种将深度学习与灵敏度分析结合的端到端方法,对深度模型的结构和激活函数进行特殊设计以适应后续灵敏度的计算,深度模型训练的同时对保存的权重信息进行反向传播计算灵敏度.在波士顿房价数据集、航迹融合数据集和G函数上的实验结果表明,所提方法相比于Sobol’法等经典方法在参数分布不均匀时准确性更高,具备更强的鲁棒性,相比于使用传统神经网络的方法准确性更高.此外,通过实验验证了基于深度学习模型的样本参数灵敏度可以优化模型的输出结果.

This paper proposes an end-to-end method that combines deep learning and sensitivity analysis,which can perform gradient back propagation calculation sensitivity on the saved weight information while training the model.The structure and activation function of the depth model are specially designed to adapt to the subsequent sensitivity calculation.The experimental results conducted on a Boston house prices dataset,a track information fusion dataset,and the G function show that the proposed method is more accurate than classical methods such as Sobol’method when the parameter distribution is uneven,and has a stronger robustness.Compared with the traditional neural network method,the accuracy of the proposed method is higher.The experiment proves that the sample parameter sensitivity obtained by the deep learning model can be used to optimize the model output.