Non linear regression and non linear approximation are widely used fordata analysis. In many applications, the aim is to build a model linkingobservations and parameters of a physical system. Two cases ofincreasing complexity have been studied: the case of deterministicinputs and noisy output data and the case of noisy input and outputdata. We present in this paper a general formulation of non linearregression using multilayered Perceptrons. Regression algorithms arederived in the three cases. In particular, a generalized learning ruleis proposed to deal with noisy input and output data. The algorithmenables not only to build an accurate model but also to re*ne thelearning data set. The algorithms are tested on two real-world problemin Geophysics. The good results suggests that multilayered Perceptronscan emmerged as an e°cient nonlinear regression model for a wide rangeof applications.