Variable structure neural networks for online identification of continuous-time dynamical systems using evolutionary artificial potential fields

A novel neural network architecture, is proposed and shown to be useful in approximating the unknown nonlinearities of dynamical systems. In the variable structure neural network, the number of basis functions can be either increased or decreased this is according to specified design strategies so that the network will not overfit or underfit the data set. Based on the Gaussian radial basis function (GRBF) variable neural network, an online identification of continuous-time dynamical systems is presented. The location of the centers of the GRBFs is analyzed using a new method inspired from evolutionary artificial potential fields method combined with a pruning algorithm. A minimal number of neuron is guaranteed by using this method. It is in noted, that both the recruitment and the pruning is made by a single neuron. By consequence, the recruitment phase does not perturb the network and the pruning dot not provoking an oscillation of the output response. The weights of neural network are adapted so that the dynamics of the system checks the imposed performances, in particular the stability of the system.