Note that this result differs from chess, for example, where speculatively
pondering is definitely better.
Why chess works out differently: the move matching rate in chess is roughly
50%, so by speculatively pondering you can increase your thinking time by
50%.
When you ponder by thinking about the opponent's play, then you gain
according
to the inverse of the branching factor. (Because you are going to throw away
one layer of the tree.) Since the branching factor in chess trees is > 2,
you will gain less than 50%. A key point in this analysis is that the move
matching rate for an N ply search is not significantly better than the move
matching rate for an N+1 ply search.
Branching factors in a UCT/RAVE tree are nominally very large, but in effect
are extraordinarily narrow. Almost all attention concentrates on a handful
of choices.
This explanation looks inexact.
The move matching will always be better than 1/branching factor. At
least if your program does not explicitely like moves your opponent
dislikes: just pick at random if necessary.
The fact that you may gain from not concentrating on one move is that it
is better to spend half the time on each move your opponent plays than
to spend all the time you can every other move. Decreasing returns.
Jonas
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
