J. Jarmulak, NeuroControl Workbench, M.Sc. thesis, Delft University of Technology, The Netherlands, August 1994.
- Neural Network training – screenshot and some discussion of training speed
Neural control offers new ways of controlling complex, nonlinear systems. The field of neural control is relatively new, and though already some practical applications of neural networks in process control exist, still much research has to be done before neural control systems will be used on a larger scale. Most of the research in this field is done using several standard software tools, which are normally used for system simulation, study of various types of neural networks, data visualization, etc. This approach costs a lot of time, is cumbersome, and is often not very flexible.
To make experiments with neural control easier to perform, a NeuroControl Workbench has been designed and implemented. The Workbench makes it possible to construct four different neural controllers: the direct-inverse, the forward-modelled inverse, the iterative inverse, and the model-based predictive. A traditional PID controller is also included for a comparison. The controllers can be used to control one of the eight predefined and two user-definable plant simulations.
An extensive use of the Workbench has shown that apart from the direct-inverse controller, all the other neural controllers achieve good results. The predictive controller is however clearly the best one. Its main advantage is, apart from the good performance, the great flexibility with which the desired controller behaviour can be specified.
The Workbench has already been used by students during a lab project.