On 04/07/2011 14:56, terry mcintyre wrote:
Chess is inherently binary - you checkmate the king or you don't. If you
don't, there are a just few simple possibilities - your king is
checkmated, or there is some form of draw.
Going back to the idea that MCTS is not perfect. It is by now
well-established that current MCTS bots - even top ones - struggle to
handle capturing races (semeai) properly. Many examples of "snatching
defeat from the jaws of victory" by way of failing to properly defend a
capturing race exist on KGS, and some have been copied to this list.
These are highly deterministic contests - exact solutions are known; one
player wins, often by exactly one move.
Imagine that the program is the winner of such a semeai, and the
hypothetical loss of the semeai would not be enough to give the human
player the win. The game, as judged by a perfect oracle, a professional
player or even a high-dan amateur, is "in the bag" for the computer. The
MCTS evaluation agrees. Based on indifference to yose (end-game) moves
which have no theoretical impact on the win, the MCTS then gives away a
few points to the human player.
After grabbing a few yose points, the balance favors the human, who then
exploits a known weakness of MCTS programs, takes one of the semeai
liberties, and the program - due to a known and exploitable weakness -
plays some meaningless move in the middle of its own territory.
At this instant, a divine oracle or human pro or even a human kyu-level
player would recognize that the computer has snatched defeat from the
jaws of victory. The human wins the semeai; this, coupled with the free
points yielded in yose, gives the victory to the human. Several moves
after, the MCTS program will realize the change in fortune, and resign.
Can you give an example of a game where a program has lost, and would
not have lost without the points yielded in yose?
Nick
Rinse and repeat. Examine the losses on KGS, and you'll find more than a
few examples. Examine the cases where the human player loses on time,
and you'll find a few more.
Don repeatedly claims that retaining yose points would reduce the
winning probability. This is seldom correct. In most cases, the more
points one yields, the closer the score is to zero, and the greater the
risk of having made a mistake in reading, or making a mistake in
playing, and discovering oneself on the losing side of zero.
When we say "rich men don't pick fights," this isn't suggesting that
rich men ( people who are winning ) give away easy points. It means to
not engage in risky fights with uncertain outcomes; look for simpler
lines which are predictable and which still lead to a maximal win.
Static twiddle factors do not use information about risk versus
stability. Do current programs even know anything about risk versus
stability?
Terry McIntyre <[email protected]>
--
Nick Wedd
[email protected]
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
