This paper concerns the application of differential evolution to training radial basis function networks. The algorithm consists of initial tuning, local tuning, and global tuning. The last two tunings both use a cycle-increased searching scheme, and global tuning employs fuzzy adaptive control. The mean square error from desired to actual outputs is applied as the objective function. Four standard test functions is used for demonstration. A comparison of net performances with two approaches reported in the literature shows the resulting network performs better in terms of a lower mean square error with a smaller network.