The “Demons in Imperfect Search” post seems related to something that’s already been studied in ML: local minima. Small changes in initial conditions can result in very different local minima being converged on. And you could say that the local minima that are selected for are the ones such that a large region of the search space will converge to that minimum if initialization starts within that region.
The “Demons in Imperfect Search” post seems related to something that’s already been studied in ML: local minima. Small changes in initial conditions can result in very different local minima being converged on. And you could say that the local minima that are selected for are the ones such that a large region of the search space will converge to that minimum if initialization starts within that region.