[computer-go] Heuristics for MC/UCT with all-or-nothing payouts

[Brian Slesinsky]

With repeat-winners, if there is a move is seems flawless at first but
some flaw is eventually found, there might be a rough transition once
the flaw is identified, since there is no backup plan.  It might make
more sense to study two apparently flawless children equally until a
flaw is found in one of them, and then to look for a new backup?

Or perhaps exploring every child once is very cheap, so avoiding it
doesn't save much time?

- Brian