Neuro-Evolution of Continuous-Time Dynamic Process Controllers

Artificial neural networks are means which are, among several other approaches, effectively usable for modelling and control of non-linear dynamic systems. In case of modelling systems input and output signals are a-priori known, supervised learning methods can be used. But in case of controller design of dynamic systems the required (optimal) controller output is a-priori unknown, supervised learning cannot be used. In such case we only can define some criterion function, which represents the required control performance of the closed-loop system. We present a neuro-evolution design for control of a continuous-time controller of non-linear dynamic systems. The controller is represented by an MLP-type artificial neural network. The learning algorithm of the neural network is based on an evolutionary approach with genetic algorithm. An integral-type performance index representing control quality, which is based on closed-loop simulation, is minimised. The results are demonstrated on selected experiments with controller reference value changes as well as with noisy system outputs.