Method feasibility study: Bayesian networks

Basic principles of Bayesian networks, inference with them and discovery of Bayesian network structures are briefly introduced. Then, the applicability of these methods to the analysis of process data is addressed. The case study problems involve mining of dependencies from training data and using the discovered dependency models for prediction of quality indicator values. Prediction results are presented as diagrams and commented. The predictions achieved are promising but it seems that with the current models the prediction accuracy is not good enough for the case problem. With suitable training data, Bayesian dependency models may be discovered from the data and applied in many ways. The possibilities range from "What- If" -analysis of the effect of value changes to the probability distributions of the other variables to sequential decision making using influence diagrams. The generated models may be implemented as C programs similarly to the way tested in this case study