In identification, variable selection for a nonlinear non-parametric and high-dimensional system is the first and often one of the most difficult problems. The issue is the curse of dimensionality. This paper presents a numerically efficient algorithm of variable selection for high-dimensional nonlinear non parametric systems. It is based on averaging derivatives and relies on one dimensional estimates of the density function and its derivative. Thus, it avoids the curse of dimensionality usually encountered for high-dimensional systems. Theoretical analysis is provided and the conditions are derived for a variable to contribute or not to contribute for a large class of nonlinear systems. Further, convergence results are established. (C) 2018 Elsevier Ltd. All rights reserved.