Abstract

We present a regression technique for data driven problems based onpolynomial chaos expansion (PCE). PCE is a popular technique in the field ofuncertainty quantification (UQ), where it is typically used to replace arunnable but expensive computational model subject to random inputs with aninexpensive-to-evaluate polynomial function. The metamodel obtained enables areliable estimation of the statistics of the output, provided that a suitableprobabilistic model of the input is available. In classical machine learning (ML) regression settings, however, the systemis only known through observations of its inputs and output, and the interestlies in obtaining accurate pointwise predictions of the latter. Here, we showthat a PCE metamodel purely trained on data can yield pointwise predictionswhose accuracy is comparable to that of other ML regression models, such asneural networks and support vector machines. The comparisons are performed onbenchmark datasets available from the literature. The methodology also enablesthe quantification of the output uncertainties and is robust to noise.Furthermore, it enjoys additional desirable properties, such as goodperformance for small training sets and simplicity of construction, with onlylittle parameter tuning required. In the presence of statistically dependentinputs, we investigate two ways to build the PCE, and show through simulationsthat one approach is superior to the other in the stated settings.

 

Introduction (beta)

None

 

Conclusion (beta)

None