The doctoral dissertations of the former Helsinki University of Technology (TKK) and Aalto University Schools of Technology (CHEM, ELEC, ENG, SCI) published in electronic format are available in the electronic publications archive of Aalto University - Aaltodoc.
|
|
|
Neural Network Methods in Analysing and Modelling Time Varying Processes
Timo Koskela
Dissertation for the degree of Doctor of Science in Technology to be presented with due permission of the Department of Electrical and Communications Engineering for public examination and debate in Auditorium S2 at Helsinki University of Technology (Espoo, Finland) on the 12th of December, 2003, at 12 o'clock noon.
Dissertation in PDF format (ISBN 951-22-6818-3) [811 KB]
Dissertation is also available in print (ISBN 951-22-6817-5)
Abstract
Statistical data analysis is applied in many fields in order to gain understanding to the complex behaviour of the system or process under interest. For this goal, observations are collected from the process, and models are built in an effort to capture the essential structure from the observed data. In many applications, e.g. process control and pattern recognition, the modeled process is time-dependent, and thus modeling the temporal context is essential.
In this thesis, neural network methods in statistical data analysis and especially in temporal sequence processing (TSP) are considered. Neural networks are a class of statistical models, applicable in many tasks from data exploration to regression and classification. Neural networks suitable for TSP can model time dependent phenomena, typically by utilizing delay lines or recurrent connections within the network.
Recurrent Self-Organizing Map (RSOM) is an unsupervised neural network model capable of processing pattern sequences. The application of the RSOM with local models in temporal sequence prediction is presented. The RSOM is applied to divide the input pattern sequences into clusters, and local models are estimated corresponding to these clusters. In case studies, time series prediction problems are considered. Prediction results gained from the RSOM model show better performance than the model with conventional Self-Organizing Map. The RSOM can capture temporal context from the pattern sequence, which is useful in the presented prediction tasks.
As another application, a neural network model for optimizing a Web cache is proposed. Web caches store recently requested Web objects, and are typically shared by many clients. A caching policy decides which objects are removed when the storage space is full. In the proposed approach a model predicts the value of each cache object by utilizing features extracted from the object. Only syntactic features are used, which enables efficient estimation and application of the model. The caching policy can be optimized based on the predicted values and a cost model designed according to the objectives of the caching. In a case study, different stages and decisions made during the data analysis and model building are presented. The results gained suggest that the proposed approach is useful in the application.
Keywords: time series prediction, temporal sequence processing, neural networks, Web cache optimization, Self-Organizing Map
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
© 2003 Helsinki University of Technology
Last update 2011-05-26