Think of it this way: we're near a local optimal point using the current policies, and we're not going to dramatically improve if we just keep using a gradient

Think of it this way: we're near a local optimal point using the current policies, and we're not going to dramatically improve if we just keep using a gradient descent type algorithm. It's time to start fresh and aim for a global optimum instead of just the local one we've been living with for so long. Is it guaranteed that we'll find it? No. Is it worth it? In my opinion, yes.