Input vector determination by false nearest neighbors

In order to generate an accurate mapping, such as a Radial Basis Function (RBF) network, it is necessary to select the inputs which will best predict the dynamics seen in the output. This is not straightforward when either the underlying physics are not precisely known, significant nonlinearities are involved, or variable time delays are present in the mapping. The method of False Nearest Neighbors (FNN) may be used for determining the dimensionality of a system and for inferring the significance of terms of the input vector. Unfortunately, this technique suffers from the curse of dimensionality and can be computationally intensive. This study investigates the potential of the FNN algorithm for identifying the most useful input regressors for a known model when input data and computing time are limited and examines several application issues.