Feedforward neural network based nonlinear dynamical system function reconstruction

In this paper, we discuss the pre-processing problems on feedforward neural network based nonlinear dynamical system function reconstruction. They are focused on input units and hidden layer units determination, training samples selection. The number of input units has a great affect on function learning. In the paper the correlation dimension is applied to determinate a suitable number of input units based on Takens theorem. Experiment shows the process bring a better reconstruction performance. We also analyze the performance of a kind of pruning algorithm. The algorithm is used to obtain a certain hidden layer units number. The experiments show that the algorithm is stable regarding reconstruction performance. Lastly we present an ad-hoc algorithm to obtain an approximate uniform distribution of training samples. Experiments results show the algorithms of training samples selection brings a great improvement in learning time and generalization performance