Optimizing the GAN loss