Black box modelling of nonlinear systems from noisy input-output data, a hybrid-neural approach. Possibilities and limitations, special questions of practical applications. The subject of the research work is the development, analysis and application of such modelling procedures that are based on measurement data, and where the construction is based on the application of machine learning. The systems to be modelled belong to different classes: such as linear systems with weak non-linearities, complex (strongly) non-linear and/or time variant systems. The main characteristic of the approach to be applied is the combination of some recently developed paradigms, like neural networks, rule based systems, etc. where the different advantageous properties of the various paradigms are exploited and combined. The most important problems to be studied are: the effects of finite training data samples, the consistency of the training data, the redundancy of the training data, whether or not all data are required, the effects of missing data and effects of noisy data. In addition to the listed main problems the research aims at to develop new hybrid intelligent architectures, that can utilize different representations of knowledge, e.g. knowledge in the form of measurement data, propositional rules, exact mathematical equations. During the research emphasis will be given to the questions of efficient implementation too.