RBF networks represent a vital alternative to the widely used multilayer perceptron neural networks. In this paper we present and examine several learning methods for RBF networks and their combinations. A gradient-based learning, the three-step algorithm with unsupervised part, and an evolutionary algorithms are introduced, and their performance compared on benchmark problems from the Proben 1 database. The results show that the three-step learning is usually the fastest, while the gradient learning achieves better precision. The best results can be achieved by employing hybrid approaches that combine presented methods.