A method called constrained topological mapping (CTM) has been recently proposed for nonparametric regression analysis (V. Cherkassky and H. Lari-NaJafi, 1990). The CTM algorithm is a modification of Kohonen self-organizing maps suitable for regression problems. The authors discuss efficient software implementations of the algorithm that may be especially attractive for multivariate problems which require a large number of units in a map. The authors present experimental comparisons with alternative neural network approaches (backpropagation) and conventional approaches (projection pursuit) to regression. These comparisons demonstrate overall superiority of the proposed CTM algorithm, both in terms of prediction and computational speed.