Using Randomized Search Algorithms to Estimate HMM Learning Parameters

Abstract

Many complex problems like Speech Recognition, Bioinformatics, Climatology, Control and Communication are solved using Hidden Markov Models (HMM). Mostly, optimization problems are modeled as HMM learning problem in which HMM parameters are either maximized or minimized. In general, Baum-Welch Method (BW) is used to solve HMM learning problem giving only local maxima/minima in exponential time. In this paper, we have modeled HMM learning problem as a discrete optimization problem such that randomized search methods can be used to solve the learning problem. We have implemented Metropolis Algorithm (MA) and Simulated Annealing Algorithm (SAA) to solve the discretized HMM learning problem. A comparative study of randomized algorithms with the Baum Welch method to estimate the HMM learning parameters has been made. The Metropolis Algorithm is found to reach maxima in minimum number of transactions as compared to the Baum-Welch and Simulated Annealing Algorithms.