Development of Optimal Neural Networks for Application in Design of Experiments

Design of Experiment (DoE) is a powerful statistical methodology adopted as an upgraded level of measurement and assessment for analysis of experiments that involve multiple factors with different levels, leading to better understanding of the relationships between them and their effects on the functionality and reliability of the experiment. Due to increased complexities and size of experiments in various fields especially in engineering and science, multiple statistical techniques were studied such as optimal DoE, fractional factorial design (FD) and combinations between other different techniques to reduce cost of analysis. From the literature review, extensive combined approaches between artificial neural networks (ANN)s and DoEs were developed to take advantage of both in terms of identifying significant factors and prediction of missing and future data in different engineering and science applications. However, a research gap in exploiting power of DoEs in optimizing ANNs hyper-parameters were very limited. This work aims to construct a feed-forward neural network (FNN) to predict p-value from post-hoc analysis results. Then, utilizing FDs to identify the significant hyper-parameters affecting this FNN and work on optimizing them to increase its accuracy by reducing error.