Generative Adversarial Networks (GANs) have demonstrated great success in image generation in recent years. This thesis proposes a novel application of GANs: the generation of mathematical formulae. In this context, a formula is described as a sequence of discrete symbols that represent a mathematical relation between quantities. These can include integers, operands and constants. Due to the well known "non-differentiability" issue of GANs in discrete data generation , the standard GAN architecture cannot be used. To that end, the architectures from two language generation models are modified and evaluated: the Gumbel-Softmax GAN  and the Sequence GAN . Experimental and mathematical results demonstrate the unviability of using the Gumbel-Softmax GAN for the generation of mathematical formulae using the proposed approaches, by cause of the difference between its intended application and the one studied in this thesis. Finally, the experiments done on the Sequence GAN show highly promising results, which make this particular architecture the more suitable amongst the analysed ones.