In this paper we describe a stochastic method for global optimization based on a uniform sampling in the search domain. After a reduction of the sample, computing the distance between the remaining points and using the distribution of the kth nearest neighbour enables clusters of points to be built up, hopefully fitting the regions of attraction of significant local optima; from each of these a local search is started. The properties of the method are analysed, and detailed computational results on standard test functions are provided.