ABSTRACT
A support vector machine (SVM) learns a classification function with two target classes by solving a quadratic programming problem. This chapter reviews the theoretical foundation of SVM that leads to the formulation of a quadratic programming problem for learning a classifier. It introduces the SVM formulation for a linear classifier and a linearly separable problem, followed by the SVM formulation for a linear classifier and a nonlinearly separable problem and the SVM formulation for a nonlinear classifier and a nonlinearly separable problem based on kernel functions. If a SVM linear classifier is applied to a nonlinearly separable problem, it is expected that not every data point in the sample data set can be classified correctly by the SVM linear classifier. The learning of an artificial neural network (ANN) requires the search for weights and biases of the ANN toward the minimum of the classification error for the training data points, although the search may end up with a local minimum.
