Have not read the paper yet, but this not sound like GAN applied to agent training?