Deep Learning Based Resource Allocation: How Much Training Data is Needed?

Abstract

We consider artificial neural networks based energyefficient power control for interference networks. The influence of different training set sizes and data augmentation is evaluated. It is shown that as few as 15,000 data points obtained from 300 channel realizations are sufficient to adequately predict almost globally optimal power allocations in a 4 user network. Moreover, we observe that, especially for larger scenarios, data augmentation is essential for successful training and far outweighs the effect of increasing the training data set size.