Is the Generational Training a Good Idea?

The way it works in with generations is that you're not actually copying the weights you just went to the installation loss. It's like a KL between the action distributions of the previous agent and your agent on the new data. I see, I see. That actually makes a lot of sense. And there's some other theory for why the generational training process might work better. This is a bit wild, but it may not just be about overfitting,. But for the weight copying, what I always felt was that weight copying kind of erases information in the network that you're copying to.