-OO or any future optimization level should try the search at multiple different starting points

[lifthrasiir]

-OO currently starts a new search at the previous best parameters, but running -O2 multiple times independently might have a better result. There may be multiple reasons: linear parameters would always stuck on local minima, and SA might not be sufficient to escape from the current local minimum. For that reason it was suggested that -O3 should just run -O2 (or equivalently and more efficiently, only the simulated annealing part) 10 times and keep the best among them.