Faster SVM training via conjugate SMO

We propose an improved version of the SMO algorithm for training classification and regression SVMs, based on a Conjugate Descent procedure. This new approach only involves modest increase computational cost each iteration but, in turn, usually results substantial decrease number iterations required to converge given precision. Besides, we prove convergence iterates this as well linear rate when kernel matrix is positive definite. have implemented within LIBSVM library show experimentally that it faster many hyper-parameter configurations, being often better option than second order performing grid-search SVM tuning.