Re-estimation of Continuous Hidden Markov Model with Multiple Observation without Overflow
He Qun Qiang · Dianzi xuebao · 2000
In the training phase of HMM system,due to the cumulative production,the evaluation of the forward and backward probabilities in Baum Welch algorithm needs a large dynamic range which will exceed the precision range of essentially any machine.This can be resolved by multiplying the forward and backward probability with scaling coefficients in single observation case.In multiple observation case,generally the output probability terms are introduced for each observation,which will cause the overflow problem again.In this paper,the cause of overflow problem is studied and a revised BW algorithm specially for multiple observation is deduced by redefining the optimization object function from cumulative production to sum of logarithms.The revised HMM parameter re estimating algorithm is more stable and effective,and the overflow problem is eliminated.