Control of nonlinear continuous-time multi-input and multi-output dynamic systems using neuroevolution

We propose a neural network-based controller of nonlinear multi-input and multi-output (MIMO) continuous-time dynamic systems. The controller learns using a genetic algorithm (GA) to minimize the chosen criterion function. A new neural network (NN) architecture is proposed that uses difference- and summation-based neurons in the hidden layers. Two control approaches are compared: a decentralized approach, which assumes that individual subsystems are isolated feedback loops, and a centralized one, which assumes that the entire system is controlled as a single system including all internal interactions between subsystems. Both approaches are compared experimentally with a conventional decentralized PID controller approach. We will show that both NN approaches can learn the plant non-linearity, however, the centralized control approach can much better eliminate the mutual internal interactions among the partial subsystems of the MIMO system.