ABSTRACT
In general, the DNN in those DNN-assisted methods can be conceptually replaced by another machine learning method, such as Support Vector Machine (SVM), Random Forest (RF), and XGBoost. We first consider a two-sample hypothesis testing problem with the scale-uniform distribution. The performance of DNN is compared with SVM, RF, and XGBoost when constructing both testing statistics and critical values, or just critical values. The DNN-based method has a better controlled type I error rate with higher power and satisfactory computational efficiency. We also revisit the point estimation problem in adaptive designs in Chapter 4 and show that DNN has a more robust and better performance than other methods. For a particular problem, the choice of a specific computational method can be treated as a discrete hyperparameter to be optimized in cross-validation.
