UP - logo
Narodna in univerzitetna knjižnica, Ljubljana (NUK)
Naročanje gradiva za izposojo na dom
Naročanje gradiva za izposojo v čitalnice
Naročanje kopij člankov
Urnik dostave gradiva z oznako DS v signaturi
  • Optimiranje verjetnosti prikritih modelov Markova z uporabo hibridnih algoritmov = Optimising probability of the hidden Markov models using hybrid algorithms
    Kaiser, Janez ; Horvat, Bogomir, 1936- ; Kačič, Zdravko
    In this paper, two new hybrid algorithms for Maximum Likelihood Estimation (MLE) of Hidden Markov Models (HMM) are presented. The primary goal of their design is overcoming the locality of the ... Baum-Welch (BW) reestimation algorithm. They combine fast optimisation using the BW algorithm with the global optimisation using a genetic algorithm. The first type of the algorithm is presented in Figure 4. It consists of two stages which are run iteratively. The first stage is the local optimisation with the BW algorithm. It is run until a local maximum is reached. In the second stage, a genetic algorithm is used to escape from the local maximum. The genetic algorithm is initalised using the maximal point from the first stage. It is run until a better point than the maximal point from the first stage is reached. This point is subsequently used as the starting point for the first stage of the algorithm in the next iteration. The second algorithm, depicted in Figure 5, is similar to the first one, except that the two stages are not iterated. Also, the genetic algorithm in the second stage of the algorithm is run until the convergence is reached or the predefined number of generations is exceeded. A new crossover operator forthe used genetic algorithm is presented. It is designed to avoid the problems which arise because of the fact that the cromosome contains unnormalised parameters of the HMM. Before the parts of the cromosomes are exchanged, groups of linked parameters lying around the crossover points are normalised to avoid unnecessary disruption. Using this operator, two different cromosomes which represent the same HMM will still represent this HMM after the crossover operation. This is not necessarily the case when the usual crossover operator is used. Results using the hybrid algorithms for the training of discrete HMM for phonemes on the Slovenian SNABI database are presented. In Table 1 the values of the objective function given by the BW algorithm, first and second type of the hybrid algorithm, respectively, are shown. It can be seen that both hybrid algorithms can escape from the local maximum and produce better values of the objective function. Recognition results for phonemes on the same database are presented in Table 2. Models trained with the second type of the hybrid algorithm yield a 1.1% reduction in the error rate over the models, trained with the BW algorithm. The first type of the hybrid algorithm yields worse results than the second type even though it assures higher values of the objective function. The probable cause for this is over-training due to a small quantity of the training data.
    Vir: Elektrotehniški vestnik. - ISSN 0013-5852 (Letn. 66, št. 3, 1999, str. 206-213)
    Vrsta gradiva - članek, sestavni del
    Leto - 1999
    Jezik - slovenski
    COBISS.SI-ID - 4625174

vir: Elektrotehniški vestnik. - ISSN 0013-5852 (Letn. 66, št. 3, 1999, str. 206-213)

loading ...
loading ...
loading ...

Vnos na polico