A maximum margin discriminative learning algorithm for temporal signals

We propose a new maximum margin discriminative learning algorithm here for classification of temporal signals. It is superior to conventional HMM in the sense that it does not need prior knowledge of the data distribution. It learns the classifier by using a nonlinear discriminative procedure based on a maximum margin criterion, providing a strong generalization mechanism. This maximum margin discriminative learning method is presented together with a two-step learning algorithm. We evaluate the kernel based hidden Markov model by applying it to some simulation and real experiments. The preliminary results have shown significant improvement in classification accuracy