Rate control mechanisms for scalable video coding are known to be a complex problem, suffering from inaccurate performance due to the layered structure of scalable video. In this study, multiple machine learning methods are evaluated to design rate-complexity-quantization models for efficient rate control. Specifically, the bit rates of SVC multi-layered bit streams are predicted, based on the quantization parameters used for encoding and a number of features describing the complexity of the video content. The results indicate that general regression neural networks can be used as an alternative to the statistical models classically used for rate control. Moreover, this approach is able to capture the influence of video complexity on the bit rate of all layers at once, and offers the possibility of increasing its effectiveness as it gains experience through on-line learning. The constructed models provide a good prediction of the encoded bit rates, with a Pearson correlation well above 0.9 and an average error of about 5%. The resulting predictor can serve as basis for a more elaborate rate control system for scalable video coding.