| Online pattern recognition has two additional features when compared to offline pattern recognition. The first feature is training data arrives in a streaming fashion for a possibly infinite time. The second feature is that the process generating the training data can change over time.;A series of solutions are presented that address a subset of the proposed problems. For each proposed solution, experimental evaluation was conducted on simulated, benchmark, and/or real data and the advantages and disadvantages of each solution are discussed. As well, the contribution of the proposed solution is explained with references to existing literature.;Fully solving these four problems would allow online intelligent systems to have the same level of accuracy as offline batch systems but have a suitable computational complexity for online systems. This goal represents an asymptotic boundary as online processing will always have additional restrictions in terms of time and memory space when compared to offline processing. It is the finding of this thesis that there is significant room for improvement in the performance of online kernel machines and the methods in this thesis take some of the necessary steps towards the ideal boundary.;The primary research problem addressed is the improvement of kernel machines operating in online environments. There are four linked subproblems that are interrelated and are addressed in this thesis. First, the computational efficiency of kernel machines is important in online applications. Second, the online adaptability of kernel machines is necessary when there is a change in the nature of the input data stream Third, the estimation accuracy of kernel machines is important because of the online training methods used. Fourth, the limited memory of the online environment combined with stochastic gradient descent gives rise to truncation error in the kernel machine estimation. |
