Reliable pavement backcalculation with confidence estimation

Abstract

Conformal Prediction (CP) is a novel machine learning concept which uses past experience to determine precise levels of confidence in new predictions. Traditional machine learning algorithms, such as Neural Networks (NN), Support Vector Machines (SVM), etc. output simple, bare predictions without an indication (confidence) of how likely each prediction is of being correct. However, in real-world problem domains, it is highly desirable to have predictions complemented with a tolerance interval to assess their credibility and accuracy. CP can be thought of as a strategy built on top of a machine learning algorithm to hedge its predictions with valid confidence. In this paper, a NN Regression (NNR) model is used to derive a decision rule for the inverse prediction of non-linear pavement layer moduli from Non-Destructive Test (NDT) deflection data. A CP is then implemented for the NNR decision rule and tested on an independent data set to demonstrate its error calibration properties. It is shown that CP can be used to derive reliable pavement moduli predictions without compromising the accuracy of the NNR decision rule but with control of the risk of error.