Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
On-line adaptation using soft-computational learning methods is on the rise for use in safety-critical applications such as fault-tolerant flight control, maintenance of distributed networks, implementation of high security devices, etc. The inapplicability of traditional analysis methods is limiting the wider use of soft-computational learning methods in safety-critical applications that involve online adaptation. The focus of the research is the development of non-conventional analysis techniques for the testing, verification, validation and analysis of adaptive learning components such as the online learning neural networks.;Our research considers stability of online adaptation as a heuristic measure of correctness in the operation of the adaptive component. The approach is based on the principles of stability according to Lyapunov theory, deriving mathematical stability proofs to assure convergence in neural network learning within a bounded amount of time. The analysis is applied to online learning neural networks such as the Dynamic Cell Structures, Sigma-Pi, and Adaline. This approach is applicable for learning from stationary, non-varying data. For time-varying training data sets, we developed the online stability monitoring methodology. Stability monitors analyze the neural network's learning in real-time. ROC curves present the performance of the developed stability monitoring system as a trade off between the selectivity and sensitivity of the stability detection. Further, we derived a convergence prediction methodology that, given the amplitude of the disturbance, predicts the number of learning cycles required by the neural network to return to a stable state. Our research identifies the significance of topology preservation for a stable online adaptation. In order to improve the robustness of topology preservation, we propose a modified DCS learning algorithm.;Our dissertation offers the first known methodology for verification, validation and analysis of learning in adaptive computing applications. The developed techniques can overcome the difficulties associated with model-uncertainty in the context of assurance of adaptive systems.
Yerramalla, Sampath, "Stability monitoring and analysis of online learning neural networks" (2005). Graduate Theses, Dissertations, and Problem Reports. 2269.