Faculty Early Career Development: Resampling Approaches to Neural Model Validation

点击翻译按钮获取中文摘要

9502134 The focus of this research is the improvement of the validation of neural network models by developing neuro-adaptations and hybrids of statistical resampling methodologies for model validation. The field of statistics has matured several approaches to statistical model validation which use the available data for both model construction and validation. Resampling approaches have the drawback that they require the construction of multiple models. The construction of several models is computationally intensive, unlike their deterministic statistical counterparts. Furthermore, there are stochastic aspects to neural network model construction that are not present in statistical modeling. This research is aimed at developing neuro-adaptations of resampling approaches by investigating the trade-offs between precision, bias, and variance of the validation procedure with the computational requirements for each of the three major resampling techniques - cross validation, jackknife, and bootstrap - along with the two most commonly used neural network validation techniques - resubstitution and train-and-test. Other research issues to be investigated and resolved are alternative error metrics, initialization of neural network weights, hybrid resampling methodologies, neural network training termination criteria, and reuse of multiple neural network models through the committee network approach. The methodologies developed will be demonstrated on typical manufacturing problems such as wave soldering of circuit boards, process planning for job shop metal forming, and other applications where data is sparse. The improvement of validation techniques for neural networks is essential for furthering research and implementation of complex systems modeling and optimization. The research has the potential to impact several disciplines currently involved in neural network modeling where data is sparse and model validation is crucial. These areas include precision manufacturing, robotic control, me dical diagnosis, and defense pattern classification and control applications. The methods to be developed can be generalized to other empirical modeling techniques where the construction of the model is computational intensive.