Collection of functions to evaluate sequences, decode hidden states and estimate parameters from a single or multiple sequences of a discrete time Hidden Markov Model. The observed values can be modeled by a multinomial distribution for categorical/labeled emissions, a mixture of Gaussians for continuous data and also a mixture of Poissons for discrete values. It includes functions for random initialization, simulation, backward or forward sequence evaluation, Viterbi or forward-backward decoding and parameter estimation using an Expectation-Maximization approach.
Release v1.2 (August 2017)
-The class of HMM with observations being modelled by a Gaussian Mixture Model (GHMM) was updated to have also a multivariated version. -The emission matrix of the GHMM model was divided into two parameters: Mu and Sigma. Mu is now a 2D matrix with number of rows equal to the observation vector dimensionality and the number of columns equal to the number of hidden states. Sigma is now a 3D matrix with number of rows and columns equal to the the observation vector dimensionality and the number of slices equal to the number of hidden states.
Release v1.1 (May 2017) -Since there are different classes of HMMs and each of them with the same algorithms, a verification step was added to avoid memory leaks and variable compatibility.