Bayesian Generational Training

The third paper was Bayesian generational population based training. And this one, I think, is the most powerful by far and it's I'd reassigned to. There's this thing with PBT methods where we can't adapt the architectures because we have this way copying component. But then last summer, the paper from DeepMind came out the Oakland Learning team. They're open-ended learning these shareholder capable agents. So they actually showed that generational training is a really effective tool for learning in open-ended environments.