Created by W.Langdon from gp-bibliography.bib Revision:1.4771
Two key issues in modelling short MTS are addressed in this thesis. Firstly, the curse of many time series modelling methods, particularly those from the traditional statistical approaches, is the number of parameters that must be located in order to apply the models. The method proposed in this thesis bypasses this parametrisation problem by locating the key relationships between the multivariate time series variables using a cross-correlation search and decomposing these variables into mutually exclusive but highly interrelated subsets using evolutionary algorithms. Secondly, the short length of these time series pose significant challenges since traditional statistical methods often place constraints on the minimum number of observations in the dataset. Towards this end, an effective way of modelling this type of data has been developed based on evolutionary algorithms and the Vector Autoregressive Process, which avoids these constraints.
The work in this thesis has been extensively evaluated against both simulated and real world biomedical time series. Evaluation is performed from both a theoretical and empirical angle and the results obtained suggest the proposed methodology is highly effective. This thesis makes the following key contributions: the introduction of a rapid correlation mining method, the effective decomposition of high dimensional MTS into subgroups, and the novel modelling of short MTS datasets.",
Supervisor Xiaohui Liu",
Genetic Programming entries for Stephen Swift