I postulate a weakness in Alpha Go Zero. Go more than any other game that comes to mind is a Gestalt exercise. Patterns rule and they flummoxed earlier coding styles. There is at least one corner of the game that is not Gestalt but involves induction. The “ladder” (‘Shichou”) is where there is a sequence of ever more forced plays and the stakes grow as the process plays out. It is important to foresee this and predict the outcome at the beginning. I am not aware of neural nets groking recursion yet. I think that recursion in humans required recognition of a pattern not on the board but a pattern in the head. I suspect that neural nets in 2017 do not do this. I think that unconscious introspection is required.
I can imagine that few expert players construct a ladder for they expect their opponents to see thru them. I wonder if Alpha Go sees thru.
Perhaps Alpha Go can brute force its way thru if there is a ‘stakes’ parameter that guides depth of search. But then who taught it “depth of search”? I suppose that I must read the paper!