Show simple item record

Evolving neuro-fuzzy tools for system classification and prediction

dc.contributor.advisorWang, Wilson
dc.contributor.authorVrbanek, Josip Jr.
dc.date.accessioned2017-06-08T13:27:22Z
dc.date.available2017-06-08T13:27:22Z
dc.date.created2008
dc.date.issued2008
dc.identifier.urihttp://knowledgecommons.lakeheadu.ca/handle/2453/3908
dc.description.abstract"Classification and prediction algorithims have recently become very powerful tools to a wide array of real-world applications. Some real world applications include system condition monitoring, bioinformatics, robotics, predictive control, earthquake prediction, weather forecasting, stock market and traffic pattern prediction, just to name a few. Within this work, several novel approaches, as well as modifications to some existing approaches, are introduced in order to improve the performance of current classification and prediction paradigms. In the first section of this work, a novel weighted recurrent neuro-fuzzy inference system is introduced alongside two existing neural networks. It is found that the novel design outperforms both the existing neural networks in terms of equal-step and sequential-step inputs for time-series forecasting. The second contribution of this work is the development of a novel evolving clustering algorithim for classification and prediction. This particular algorithim starts without any priori knowledge of the distribution of the data set. The novel design is capable of revealing the true cluster configuration in a single pass of the data, estimating the location and variance of each cluster. After a rigorous performance evaluation, it is found that the novel design outperforms many existing clustering approaches including the well-known potential-based evolving Takagi-Sugeno (eTS) clustering scheme. The third and fourth contributions of this work are the development of a second novel clustering technique and a novel hybrid training technique. The clustering technique is a combination of the aforementioned scheme and the potential-based technique. The new training algorithm is a combination of the decoupled-extended Kalman filter (for the backward pass) and the recursive least-sequares estimate (for the forward pass). It is found that the novel clustering technique outperforms many available clustering techniques. Also, the novel training algorithm is proven to outperform most existing training techniques."--Abstract
dc.language.isoen_US
dc.subjectNeural networks (Computer science)
dc.subjectFuzzy logic
dc.titleEvolving neuro-fuzzy tools for system classification and prediction
dc.typeThesis
etd.degree.nameMaster of Science
etd.degree.levelMaster
etd.degree.disciplineEngineering
etd.degree.grantorLakehead University


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record