The Optimization of Hidden Markova Model Training Procedure Using Clonal Selection Algorithm

Hidden Markova model (HMM) is an important modeling tool for the dynamic time series. An intelligent recognition system based on hidden Markova model was presented in this paper and the optimization of HMM training procedure was researched. Because the high-dimension of HMM parameter space and the local convergence of Baum-Welch algorithm, clonal selection algorithm (CSA), as a typical algorithm of artificial immune system (AIS), was adopted to increase the efficiency of HMM training. Based on the further research of HMM training procedure using CSA, best set clonal selection algorithm (BCSA) was proposed in order to increase the diversity of algorithm. The experiments using pump signals showed that the AIS aided training procedure especially the BCSA aided algorithm, was feasible and effective. The average log-likelihood CSA aided algorithm in experiments was –9.43, which was greater than the –9.64 of normal algorithm. But the training time of CSA aided algorithm was unstable. BCSA aided algorithm not only increased the average log-likelihood of 0.1 but reduced 30.1% average training time than CSA. The BCSA aided algorithm significantly improved the efficiency of HMM training.