Second-order optimization of variable projection SVM models and road abnormality detection
The study introduces a new second-order optimization framework to train Variable Projection Support Vector Machines (VP-SVMs). This method uses second-order derivatives of the objective function to efficiently find optimal parameters for VP-SVM models. The technique was successfully applied to recognize road surface abnormalities using 1D signals from tire sensors, showing that the second-order trust region (TR) algorithm significantly outperforms standard gradient descent in terms of accuracy and training speed.