Evaluation of tree based statistical learning supporting the bespoke manufacturing of low volume, super luxury cars

Manufacturing low-volume bespoke, super luxury cars is challenging. Many difficulties arise; the relatively low volume of data, the variation of the vehicles hiding patterns, the human builders lowering the consistency, for example. Yet all of these problems must be managed to create perfection. Solving these problems requires careful data selection, choosing features that make an impact; the specification of each vehicle, the production associates who built it, amongst others. Furthermore, to predict how a vehicle will perform, given the information known about it, a suitable machine learning method needs selecting. In this thesis, the process of choosing the most appropriate problem is discussed: predicting how vehicles will perform in their final inspection. Selecting the best features and engineering them into a suitable form. Established machine learning methods are discussed: random forest, XGBoost, optimal tree ensembles, neural networks and semi-supervised generative adversarial networks, as well as some novel methods, utilising P-Value adjusted decision trees in the optimal tree ensemble framework, residual boosting and conditional inference trees. These methods are compared to determine the most effective approach for predicting final inspection outcomes. The results demonstrate that the random forest model achieved the highest predictive performance with a precision of 66%, outperforming a novel method of PASTE-boost with a precision of 52%. The findings show that the final inspection result is predictable but improvements may require more observations or more influential predictors to allow for a more reliable prediction. This thesis demonstrates the potential and limitations of machine learning approaches for supporting quality assessment for low-volume manufacturing processes.