ABSTRACT
Maximum unit weight (MUW) and optimum moisture content (OMC) are commonly de-rived from laboratory testing. Alternatively, these parameters can be ascertained using regression models, artificial neural network (ANN) and machine learning (ML). While laboratory testing is still the most accurate method, an alternative method can be helpful for the approximation of these values. In this study, one of such methods, i.e. neural architecture search (NAS) techniques to define ANN for prediction, has been explored. Two NAS methodologies, NAS with hyperband (HB) (Model 1) and NAS with Bayesian optimization (BO) (Model 2) are implemented and compared based on the evaluation criteria. The results demonstrate that BO provides the best generalization performance and accurate results while HB optimization could not generalize well due to overfitting issues. This study highlights the effectiveness of NAS-based ANN models in designing fully automated architecture for prediction in geotechnical engineering.
