Machine learning and its application to process control

Methods of Machine Learning - a main topic of AI- research-are to day in a state to get major industrial applications. They can be especially applied to the analysis, diagnosis and control of complicated processes for example in production automation, ecology or economy which cannot (or only partly) be modelled by conventional mathematical methods. By an automatic analysis of a set of examples (training set) of measured input-output behaviour of the process learning algorithms can find out important causal relationships between process variables and construct classification procedure for the detection of dangerous or unwanted process states. Additionally, by classification learning partitions of the process states are induced which may be used for the construction of optimal control trajectories in process planning or even in real time process control. Two main possibilities of application of classification learning to process diagnosis will be considered: decision trees, and artificia l neuronal nets of the MADALINE-type. Now modifications and extensions are elaborated within the ESPRIT-project StatLog.