Need ideas on how to improve GANs.

Want to create a decent game based on GANs. However, GANs are really shit at the moment.

Has anyone here got any theories on GANs and how to improve them.

Dominate idea now

Generator -> Discrimination (detector)

Start with both being untrained, then progressively train both the generator and discriminator. The discriminator does back prop on the Generator.

I was thinking maybe ideas from dreaming can help. Saying that I was looking at sleep in general. I was thinking a awake would correspond to a reinforcement learning algorithm -> sleep would correspond to generator / discriminator idea.