Multifidelity control variates for supervised learning
Elizabeth Qian
Georgia Institute of Technology
Regression methods are popularly used to solve supervised learning problems in order to learn efficient surrogate models from data. In many of these methods, Monte Carlo estimators are used to estimate errors or parameters from data. These methods therefore suffer in scarce data settings, where the limited number of available samples leads to high variance and high error in the learned model. This is a common scenario in engineering and science application, where the high cost to obtain data via experiment or simulation leads to extreme data scarcity. Multifidelity control variate methods are a class of methods that exploit lower-fidelity models to obtain additional low-fidelity data that can be used to reduce the variance of Monte Carlo estimators. In this talk, we will introduce and discuss the use of multifidelity control variate methods in regression problems for supervised learning, and demonstrate numerically that this leads to learned models with improved accuracy and robustness in scarce data settings.