MOTIVATION Most "real world" systems relevant from a control perspective exhibit nonlinear behavior and are often difficult/unrealistic to model using laws of physics, chemestry, etc.. The neural network has shown to be a promising technique to overcome this problem as it is useful for building good models from measured data. If a neural network model is available, different approaches are possible when designing a control system. It is possible to use a number of conventional nonlinear design techniques. E.g., feedback linearization, Generalized Predictive Control (GPC), or linearization of the model followed by a (time varying) linear design. Another approach is to use a neural network as the controller as well; e.g., direct inverse control or Internal Model Control (IMC).      ACTIVITIES The following issues are addressed in our research: Network Training: Development of effective and easy-to-use training methods and techniques for finding the "optimal" network architecure. Modelling: Different input-output model structures as well as state space model structures are considered. Deterministic as well as stochastic. Control Design: Issues like system behavior, available hardware, sampling frequency, robustness, and desired tuning properties, influence the choice of controller design. A number of different designs are developed. Experimental Verification: Implementation of various control strategies in a real time environment and testing on different lab. set ups. We have developed two toolsets for MATLAB. For network training and system identification we have implemented the toolbox NNSYSID, and for design and simulation of different control system concepts we have developed a collection of tools called NNCTRL.   MORE INFORMATION For more information, please contact one of the following: Magnus Nørgaard, Ole Ravn, or Paul H. Sørensen, Much of this work is done in collaboration with Department of Mathematical Modelling (IMM)