very interesting responses, and yet no optimal strategy.
A "greedy algorithm" by making locally optimal WL selection is probably not a bad start.
Then a bit of exploration, which comes at some cost.