*2.5. Classification*

Accurate classification of gait patterns is the ultimate goal of the user intent recognition (UIR) system. For this purpose, we assess various well-known linear and nonlinear classification techniques, including linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), support vector machine (SVM), decision tree (DT), and multi-layer perceptron (MLP) classifiers.

LDA and QDA classifiers do not require time-consuming iterations for training. In fact, the parameters of these classifiers are directly obtained from the training data. Although these classifiers are fast in terms of training, they are not as flexible as nonlinear classifiers such as SVM, DT, and MLP. These classifiers solve an optimization problem that minimize the classification error. In most cases, it is difficult to find optimization solutions in closed form, so we use either gradient-based optimization algorithms such as steepest descent, or evolutionary algorithms (EAs).

We use one-against-one approach to implement multi-class SVM, and we also evaluate the performance of different kernels, such as linear and RBF. We tune the parameters of the SVM kernels to achieve the best classification performance. To increase the accuracy of the MLP network, we perform a grid search of the number of hidden nodes *p* from the set {3, 4, 5, 6, 8, 10, 15, 20}, and we measure the mean classification error using five-fold cross validation (CV). Then, we choose *p* to obtain a trade-off between classification accuracy and classifier complexity. An MLP with small *p* may not result in the desired accuracy, but an MLP with large *p* may tend to memorize the noise in the training set and lead to overfitting and poor generalization. In addition, we use Wilcoxon signed-rank tests to statistically compare the classification methods.
