In this paper we propose a machine learning approach based on generative adversarial networks (GAN) for synthesizing novel radar waveforms with a desirable Ambiguity Function (AF) shape and constant modulus property. There are only a limited number of code sequences of a certain length in many widely used radar code families, which may be a drawback in modern radar applications. Hence, there is a need to generate new waveforms for future MIMO, multifunction, and cognitive radars. In such systems multiple waveforms are launched simultaneously in order to deal with low observable targets or a large number of small targets. Additionally, the ability to generate new waveforms at will makes it more difficult for an adversary to recognize or detect that it is illuminated by a radar. A Wasserstein GAN (WGAN) structure is developed for complex-valued input data. The model is trained using Frank and Oppermann codes with good autocorrelation and crosscorrelation properties. The synthesized novel waveforms have an almost identical AFs to those of the training data, as well as a low cross-correlation relative to the codes in the training set. Additionally, the constant modulus property facilitates the efficient use of amplifiers.