The initial set of weights to be used in supervised learning for multilayer neural networks has a strong influence in the learning speed and in the quality of the solution obtained after convergence. An inadequate initial choice of the weight values may cause the training process to get stuck in a poor local minimum or to face abnormal numerical problems. Nowadays, there are several proposed techniques that try to avoid both local minima and numerical instability, only by means of a proper definition of the initial set of weights. The focus of this paper is in the application of genetic algorithms (GA) as a tool to analyze the space of weights, in order to achieve good initial conditions for supervised learning. GA's almost-global sampling compliments connectionist local search techniques well, and allows us to find some very important characteristics in the initial set of weights for multilayer networks. The results presented are compared, for a set of benchmarks, with that produced by other approaches found in the literature.