Generalization, Hyperparameter Optimization, and Robustness of Machine-Learning Models for Asphalt Mixture Dynamic Modulus

Abstract

Machine-learning prediction of asphalt mixture dynamic modulus can reduce dependence on repeated laboratory testing, but small datasets make hyperparameter selection and generalization especially important. This article examines the optimization and robustness behavior of six learning algorithms—BP, LR, DT, KNN, RF, and SVM—using the seven predictor variables adopted from the Witczak 1-40D framework. The source dataset contains 144 laboratory observations from 16 mixtures, with 100 observations used for training and 44 held out for testing. Ten-fold cross-validation was coupled with a modified beetle antennae search (MBAS) algorithm, which introduces Levy-flight and self-adaptive inertia mechanisms to improve search behavior. The best-fold tuning RMSE ranged from 175 MPa for RF to 1,000 MPa for LR. Independent-test results further revealed major differences in generalization: RF achieved 460.774 MPa test RMSE and R = 0.9921, while DT deteriorated to 1,699.698 MPa and R = 0.8719. A derived train–test error-gap indicator shows the largest instability for DT and comparatively stable behavior for KNN and LR, while RF uniquely combined very low test error with high correlation. The study argues that optimizer performance should be evaluated jointly with model stability and held-out prediction rather than convergence speed alone.