Parametrized neurocontrollers

Neural network controllers developed with existing approaches are optimized for specific process models and control criteria. The neural networks must be individually trained for different processes and retrained for process variations and changes in control objectives. The concept of "parametrized neurocontrollers" (PNCs)-neurocontrollers with inputs that are used to adjust control system performance and to provide information about the process dynamics - is introduced. PNCs are optimized in simulation over spaces of process models and performance criteria; application-specific training is not needed. The authors discuss neurocontrollers can be optimized for robust performance to be, by design, relatively intolerant of process/model mismatch. A simple illustration of PNC concept is described.<<ETX>>