Presenter

Miss Tingting Liu - ETRO, Vrije Universiteit Brussel [Email]

Abstract

Hidden Markov Models (HMMs) are models based on unknown, hidden states for modeling dynamic systems. The main goal of my PhD is to increase our understanding of HMMs, to provide insight into the essential features of HMMs instead of using them as black boxes. Although the models are quite simple, not all questions regarding identifiability, learnability and applicability are yet answered.
For understanding the identifiability of HMMs, we studied what makes a hidden state 'important', in the sense that it has a large impact on the likelihood of observation sequences. It turns out that persistent and transient, cyclic states make an HMM highly specific. The specificity of an HMM is defined as the likelihood difference with the best HMM having one hidden state less. Novel conditions for model equivalence are proposed which can be used to check model minimality and model specificity. The proposed approach provides an instructive understanding on the relations between model parameters and model identifiability.
When considering the learning of HMMs, the predominant learning algorithm is still the Baum-Welch algorithm. It is an iterative learning procedure starting with a predefined size of state spaces and randomly-chosen initial parameters. However, wrongly chosen initial parameters may cause the risk of falling into a local optimum and a low convergence speed. Based on the understanding of what makes a state strong, we developed an approximate identification method to estimate the number of states and model parameters directly from the input data. This resulted in an improved model initialization approach for the learning algorithm. Experimental results show a much faster convergence and accuracy, especially for highly specific models. As a result, the learning speed is significantly increased with comparable accuracy to the traditional Baum-Welch algorithm. We also noticed that the higher its specificity, the easier an HMM can be learned.
The HMMs were applied on both synthetic and real industrial data for machine maintenance. The latter is from the project of Prognostics for Optimal Maintenance (POM) funded by the Flemish government in Belgium. For practical applications, a methodology for robustness and reliability was developed to ensure that the models, with certain confidence intervals, perform effectively under various changing conditions. Different model classes were compared to HMMs such as Multinomial models, Markov models and Hidden semi-Markov models.

Short CV

Master in Applied Computer Science, VUB, 2009
Master of Science in Information Management, KULeuven, 2012