Open Collections

Application of learning theory in neural modeling of dynamic systems

University of British Columbia

Neural networks have been successfully used to model a number of complex nonlinear systems. Although neural networks can create successful models of some nonlinear systems, they are known to overfit the data in some other applications. Therefore, in order to use neural networks reliably, it is necessary to explore the conditions under which neural models perform equally well on the testing and training data sets. This calls for the design of the neural models that create a balance between the testing and training performances. The newly introduced Probably Approximately Correct (PAC) learning theory addresses the issue of testing-training balance. However, conventional PAC learning only allows static modeling and cannot be applied to dynamic models. In this thesis, PAC learning is extended to more general learning schemes that handle dynamic modeling tasks. The resulting PAC paradigms are then applied to assess the learning properties of several families of dynamic neural networks, including Radial Bases Functions Networks, single-hidden-layer Sigmoid Neural Networks, and Volterra Networks. Another concern with the use of neural networks for some dynamic modeling tasks is the issue of stochastic stability. Little is known about the stochastic stability of many neural models used in practical applications. The lack of knowledge over the stability of neural models further limits the use of such models. In this thesis, sufficient conditions for stochastic stability of different families of neural networks are presented, which address the above mentioned concern. Based on the resulting learning frameworks, evolutionary algorithms are then presented that search for a suitable suitable dynamic neural modeling which perform equally well on testing and training data. The evolutionary algorithms are then used for modeling of two applications. The first application deals with next-scan-estimation of a two-dimensional paper basis weight measurement on a paper machine. In the second application, a neural model for a neuromuscular blockade system is developed. The results indicate that accurate and reliable dynamic neural models can be obtained, provided that the learning complexity of such models are controlled during the training procedure.

Thesis/Dissertation

application/pdf

2009-07-27

For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.

http://hdl.handle.net/2429/11287

Electrical and Computer Engineering

University of British Columbia

UBCV

DSpace