Quoting Darren Cook <[email protected]>:
When the programs select the best move to play it also knows what move
is the *most likely best move* for the opponent given search to that point.
With speculative pondering all pondering goes into searching the reply
to that *most likely best* move.
If the opponent plays a different move, this move will on average be a
worse move than the expected one. Although there was no gain in
pondering, my program has a better than expected position (on average).
I think this is too optimistic: it is surprisingly rare that any given
board position has just one best move. E.g. I've analyzed complex 9x9
middle games (*) where the played move lead to a 0.5pt loss; in trying
to find a win I've discovered 3 or more other moves that also lead to a
0.5pt loss. (After all, if this wasn't the case MCTS wouldn't work so well.)
If we assume your opponent is roughly the same strength as you, then
when they play an unexpected one the chances are they've not made a bad
move, just an alternative good one.
You are right of course, but I think in practice MC-programs tends to
discover the move that makes search most difficult for the opponent.
(Not everytime of course but on average). This is not the same thing
as playing the theoretically best move. So even if many move are
equal, there can still be a move that has better practical winning
chances. And the best guess is the move that is best in the search
sofar.
A separate issue which probably makes things complicated is when to
stop searching. Should the program stop searching when it know for
sure what to play? Or it should it search until it knows what move to
ponder? And how does the ponder strategy interact with opponent
programs. Good results on CGOS may just mean that the program can
exploit one or more programs with a bad time allocation.
Best
Magnus
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
