Hidden Markov models (HMMs) are models that can capture complex behavior over time. They achieve this by using graphical representations of a latent part (the states of a process) and an observable part (emissions). However, in many cases the state space can become rather large, leading to decreased problem insight and tendency of overfitting. One model that attempts to decrease the state space is the Asymmetric HMM (HMM-A) which introduces the notion of asymmetry, where the structure of the emission space can differ per state. This is in contrast to the homogeneity of conventional HMMs. HMM-As achieve state space reduction by allowing the emission space to be more expressive. However, while this is an improvement, we have investigated if the state space may be further reduced while retainingmodel quality and improve problem insight by introducing auto-regression (AR) on the emission space. In this thesis we have defined a new model (HMM-AA) which utilizes AR on the emission space. Furthermore, we provide empirical results based on experiments and a case study to find evidence that this model can achieve the above mentioned goal. Although the model proves more complex to interpret, it can more accurately represent processes and gains problem insight which the HMM-A cannot achieve.
- Master Software Engineering