Neural networks as process controllers-optimization aspects

Neural networks are now being extensively used as feedback controllers. The authors overview the basic approaches to neurocontroller development and concentrate their attention on the "model-based neurocontrol design" approach. Controller design is viewed as an optimization problem, and a basic distinction is made between gradient-based and nongradient-based algorithms. The former impose constraints on the design problem in order to facilitate the computational aspects of the optimization, whereas nongradient-based optimization allows for general problem formulations but at significant computational cost. The "Parametrized Neurocontroller" concept is discussed to motivate the need for nongradient-based optimization and an evolutionary optimization algorithm is presented.