Life In 19x19 http://www.lifein19x19.com/ |
|
Opening problems for AI: Problem 26 http://www.lifein19x19.com/viewtopic.php?f=15&t=17629 |
Page 1 of 1 |
Author: | Bill Spight [ Fri Jul 10, 2020 2:07 am ] |
Post subject: | Opening problems for AI: Problem 26 |
Black's textbook play loses 28½% to par, according to Elf, assuming komi, as always. Black to play. There are lots of possible plays out there. Enjoy! |
Author: | RobertJasiek [ Fri Jul 10, 2020 3:43 am ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Author: | Harleqin [ Fri Jul 10, 2020 3:57 pm ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Author: | pwaldron [ Fri Jul 10, 2020 5:07 pm ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Author: | John Fairbairn [ Fri Jul 10, 2020 5:36 pm ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Author: | Bill Spight [ Fri Jul 10, 2020 7:31 pm ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
A brief word on the extreme differences in Elf's estimated winrates. In theory they are estimates of the probabilities of winning from particular positions in self play games. But, AFAICT, nobody has ever checked those estimates to see how accurate they are. They are only indirectly validated by the overall strength of the program. None of the bots' winrate estimates apply to human play, although there is a tendency to interpret them that way. Nor do they apply between bots. If Elf estimates Black's winrate in a certain position as 90%, Leela Zero might estimate it as 72%. That does not mean that they disagree, because they are estimating different things. Elf's estimate applies, in theory, anyway, to Elf's self play and Leela Zero's estimate applies, in theory, to Leela Zero's self play. They could both be right. My guess is that they are both wrong. But that kind of difference between the bots is typical. Elf makes more extreme estimates than other top bots, and the differences in its estimates for different plays tend to be greater, as well. It is important to keep in mind that Elf's estimates apply only to its own self play, not to any other bot, and certainly not to human vs. human play. Some years ago I came up with a rule of thumb for estimated winrate differences for Leela 11, that their margin of error was about 3%. Leela Zero is far stronger, but that does not mean that its margin of error is smaller. It may even be larger, quien sabe? My rule of thumb for Elf was 5%, but I have recently raised it to 6%, but maybe that's too high. Anyway. a difference of 28% means that the play with the lower evaluation is almost certainly worse than the other play, and hence is a mistake. Most likely a blunder. If Leela Zero calculates a difference of 10%, that does not mean that that smaller difference does not indicate a blunder. |
Author: | RobertJasiek [ Fri Jul 10, 2020 9:18 pm ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Harleqin wrote: Harleqin, how do you want to follow-up your sequence? |
Author: | Bill Spight [ Sat Jul 11, 2020 10:50 am ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Author: | Knotwilg [ Sat Jul 11, 2020 12:12 pm ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Without looking at the other answers: |
Author: | Uberdude [ Sun Jul 12, 2020 3:43 am ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Elf wanting to defend solidly as black, and push and peep as white, is another example of how much bots hate letting the opponent make profitable exchanges in sente, so it's worth playing the reverse sente to avoid that. And bonus if your reverse sente actually had a big follow-up too (the corner atari) such that it might even be sente, in which case it's a double sente and those are even better. |
Author: | Bill Spight [ Sun Jul 12, 2020 4:06 am ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Uberdude wrote: Elf wanting to defend solidly as black, and push and peep as white, is another example of how much bots hate letting the opponent make profitable exchanges in sente, so it's worth playing the reverse sente to avoid that. And bonus if your reverse sente actually had a big follow-up too (the corner atari) such that it might even be sente, in which case it's a double sente and those are even better. I was struck how it seems that the approach to the top right corner made the peep sente. |
Author: | dhu163 [ Fri Jun 04, 2021 10:30 am ] |
Post subject: | Re: Opening problems for AI: Problem 26 |
Page 1 of 1 | All times are UTC - 8 hours [ DST ] |
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |