Hidden Markov Models Training Using Population-based Metaheuristics

In this chapter, we consider the issue of Hidden Markov Model (HMM) training. First, HMMs are introduced and then we focus on the particular HMM training problem. We emphasize the difficulty of this problem and present various criteria that can be considered. Many different adaptations of metaheuristics have been used but, until now, few extensive comparisons have been performed for this problem. We propose to compare three population-based metaheuristics (genetic algorithm, ant algorithm and particle swarm optimization) with and without the help of a local optimizer. These algorithms make use of solutions that can be explored in three different kinds of search space (a constrained space, a discrete space and a vector space). We study these algorithms from both a theoretical and an experimental perspective: parameter settings are fully studied on a reduced set of data and the performances of algorithms are compared on different sets of real data.