The most difficult problem ever. Igo Hatsuyoron 120
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
In May 2024, I started training a 28b network using the training data previously generated by the 60b network.
In July 2024, the 28b network had caught up (after a good 140 generations), so the self-play with the 60b network was terminated and continued with the 40b network.
In October 2024, I then experienced massive hardware problems, which seriously slowed down further development.
So I now concentrated the self-play training on the non-Main-Semeai variations, with a few more of these variations emerging from time to time in the games between Karl's 40b network and my 28b network.
For some of these non-Main-Semeai variations, it seems to have been consolidated in the meantime that they represent valid solutions to the problem. I will publish details on this in subsequent posts.
In July 2024, the 28b network had caught up (after a good 140 generations), so the self-play with the 60b network was terminated and continued with the 40b network.
In October 2024, I then experienced massive hardware problems, which seriously slowed down further development.
So I now concentrated the self-play training on the non-Main-Semeai variations, with a few more of these variations emerging from time to time in the games between Karl's 40b network and my 28b network.
For some of these non-Main-Semeai variations, it seems to have been consolidated in the meantime that they represent valid solutions to the problem. I will publish details on this in subsequent posts.
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
088094_F18A6
Black played the Gezumi
at the latest possible moment.
White did not follow the Main Semeai by answering locally at a, but immobilized Black's Nakade stones at the left edge by blocking with
, instead, usually inserting the sente sequence in the upper left corner after Yamada Shinji's Tsuke of
before.
The diagrams below show the most played lines in the games between the 28b and the 40b net.
Black = 28b / White = 40b Black = 40b / White = 28b
Those who want to study this non-Main-Semeai variation any further on their own will find below compilations of all 28b/40b games ending in Jigo.
The comments indicate the number of games in which the respective variation was selected.
E.g. "A=19 // B=264" == Variation "A" was played in 19 games, variation "B" in 264 games.
(The labelling is designed for SGF-editors that label variations with capitals "A", "B", "C", ... Here, in the forum's SGF-browser's diagrams, you will have to take "1" as "A", "2" as "B, "3" as "C", ...)
Black played the Gezumi
White did not follow the Main Semeai by answering locally at a, but immobilized Black's Nakade stones at the left edge by blocking with
The diagrams below show the most played lines in the games between the 28b and the 40b net.
Black = 28b / White = 40b Black = 40b / White = 28b
Those who want to study this non-Main-Semeai variation any further on their own will find below compilations of all 28b/40b games ending in Jigo.
The comments indicate the number of games in which the respective variation was selected.
E.g. "A=19 // B=264" == Variation "A" was played in 19 games, variation "B" in 264 games.
(The labelling is designed for SGF-editors that label variations with capitals "A", "B", "C", ... Here, in the forum's SGF-browser's diagrams, you will have to take "1" as "A", "2" as "B, "3" as "C", ...)
- Attachments
-
- IgoHatsuyoron120_088094_F18A6_28bWhite#199.sgf
- (120.97 KiB) Downloaded 839 times
-
- IgoHatsuyoron120_088094_F18A6_28bBlack#199.sgf
- (121.37 KiB) Downloaded 833 times
Last edited by Cassandra on Sat Feb 08, 2025 2:18 am, edited 1 time in total.
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
083095_A6N13
This is one of the "newer" non-Main-Semeai variations.
In the upper right, the Guzumi-Exchange
/
has already taken place, and Black also pushed into White's corner with
. White, however, fell back to
, instead of blocking at a. Black left the "usual" continuation for further reducing White's corner from Black a to White f for later.
At the left edge, Black started the Nakade Sequence with
at the bottom, instead of playing at g at the top.
leaves the Nakade-Sequence and starts the further reduction of White's upper right corner.
White, however, plays Tenuki with
, in order to immobilize Black's Nakade stones at the left edge.
Black is forced to secure one eye for his large group by capturing one White stone with
at the upper edge, ...
... and the second one by arresting another White stone with
in the centre.
The diagrams below show the most played lines in the games between the 28b and the 40b net.
Black = 28b / White = 40b Black = 40b / White = 28b
Those who want to study this non-Main-Semeai variation any further on their own will find below compilations of all 28b/40b games ending in Jigo.
The comments indicate the number of games in which the respective variation was selected.
E.g. "A=138 // B=5" == Variation "A" was played in 138 games, variation "B" in 5 games.
(The labelling is designed for SGF-editors that label variations with capitals "A", "B", "C", ... Here, in the forum's SGF-browser's diagrams, you will have to take "1" as "A", "2" as "B, "3" as "C", ...)
This is one of the "newer" non-Main-Semeai variations.
In the upper right, the Guzumi-Exchange
At the left edge, Black started the Nakade Sequence with
White, however, plays Tenuki with
Black is forced to secure one eye for his large group by capturing one White stone with
... and the second one by arresting another White stone with
The diagrams below show the most played lines in the games between the 28b and the 40b net.
Black = 28b / White = 40b Black = 40b / White = 28b
Those who want to study this non-Main-Semeai variation any further on their own will find below compilations of all 28b/40b games ending in Jigo.
The comments indicate the number of games in which the respective variation was selected.
E.g. "A=138 // B=5" == Variation "A" was played in 138 games, variation "B" in 5 games.
(The labelling is designed for SGF-editors that label variations with capitals "A", "B", "C", ... Here, in the forum's SGF-browser's diagrams, you will have to take "1" as "A", "2" as "B, "3" as "C", ...)
- Attachments
-
- IgoHatsuyoron120_083095_A6N13_28bWhite#199.sgf
- (58.09 KiB) Downloaded 748 times
-
- IgoHatsuyoron120_083095_A6N13_28bBlack#199.sgf
- (48.01 KiB) Downloaded 730 times
Last edited by Cassandra on Sat Feb 08, 2025 2:17 am, edited 1 time in total.
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
-
RobertJasiek
- Judan
- Posts: 6273
- Joined: Tue Apr 27, 2010 8:54 pm
- GD Posts: 0
- Been thanked: 797 times
- Contact:
Re: The most difficult problem ever. Igo Hatsuyoron 120
Do I understand it correctly that the problem is still far from being solved?
What score does current research suggest and what was the originally believed score?
What amount of total time have you spent on the problem thus far? How much all researchers together?
What hardware, libraries and software have you been using for Katago? What is your current minimum number of visits per second? How much time have you let Katago search?
What score does current research suggest and what was the originally believed score?
What amount of total time have you spent on the problem thus far? How much all researchers together?
What hardware, libraries and software have you been using for Katago? What is your current minimum number of visits per second? How much time have you let Katago search?
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
What do you understand by "solved"?RobertJasiek wrote:Do I understand it correctly that the problem is still far from being solved?
Karl's training of his 40b network and my training of the (initially) 60b network (later 28b) followed completely different philosophies.
Nevertheless, all came to an identical result: the Main Semeai is favoured, the best result for both sides is "White + 2" (with Komi = 1).
In contrast to Karl, I was unable to find access to the “hintpos” support of the training, with which KataGo is able to detect its own incorrect evaluations.
As a “substitute” I used the 60b/40b games (later 28b/40b), in which 40b could find the mistakes of 60b (28b).
As an actually unintended side effect, non-Main-Semeai variations occasionally appeared in these games (usually initiated by 60b/28b), which (if played correctly) also end with "White + 2" (with Komi = 1).
Are all basic non-Main-Semeai lines already known? No idea (probably not).
In addition, many of the known ones (especially those that start early) have yet to undergo a (very lengthy) evaluation.
See above: "White + 2" (with Komi = 1).RobertJasiek wrote:What score does current research suggest and what was the originally believed score?
The "originally believed score" is hidden in the mists of history.
Problem 120 is one of only two problems of Igo Hatsuyôron for which Inoue Dôsetsu Inseki did NOT give a result. It simply says “黒先” = “Black first”.
We can only conclude from the (now known) title "孔明仲達勢" = "Zhuge Liang and Sima Yi" of the problem by consulting the history of ancient China that White is expected to maintain the upper hand.
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
-
RobertJasiek
- Judan
- Posts: 6273
- Joined: Tue Apr 27, 2010 8:54 pm
- GD Posts: 0
- Been thanked: 797 times
- Contact:
Re: The most difficult problem ever. Igo Hatsuyoron 120
A solution to a problem (and a presumed ruleset) is established by mathematical proof.
Different degrees of solutions can be studied:
weakest: proof of the correct score, without showing any complete sequence
weak: proof for at least one stated sequence to construct the correct score
strong: for all correct sequences to the correct score
Different degrees of solutions can be studied:
weakest: proof of the correct score, without showing any complete sequence
weak: proof for at least one stated sequence to construct the correct score
strong: for all correct sequences to the correct score
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
What task do you have in mind here?RobertJasiek wrote:What hardware, libraries and software have you been using for Katago? What is your current minimum number of visits per second? How much time have you let Katago search?
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
-
RobertJasiek
- Judan
- Posts: 6273
- Joined: Tue Apr 27, 2010 8:54 pm
- GD Posts: 0
- Been thanked: 797 times
- Contact:
Re: The most difficult problem ever. Igo Hatsuyoron 120
Anything when you search for something larger than one local LD problem. E.g., when you do a whole board search. In particular then minimum number of visits per second.Cassandra wrote:What task do you have in mind here?RobertJasiek wrote:What hardware, libraries and software have you been using for Katago? What is your current minimum number of visits per second? How much time have you let Katago search?
Libraries: versions of CUDA, CuDNN, TensorRT.
Software: versions of KataGo and what model net, which Go GUI for calling KataGo.
Hardware: uh, which GPU(s), to start with etc.
Time: for all your H120 studies since you started many years ago.
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
For the games between 60b/28b and 40b are combined:RobertJasiek wrote:Anything when you search for something larger than one local LD problem. E.g., when you do a whole board search. In particular then minimum number of visits per second.
- Colours: 40b = Black / 40b = White
- Thinking time: 60s/45s and 30s / 10s/7.5s and 5s / 2s/1.5s and 1s / 100/100 and 100 visits (seconds corresponding to the recpective net's size)
- "normal" PLAY / PLAY with additional parameters once recommemded by David for ANALYSIS, resulting in braoader, but less deaper search
"60s/45s and 30s" resulted in 10.000.000 down to x0.000 root visits per move.
"Root visits per move" depends on the net, the position, and the familiarity of the net with the position, so "visits per second" does not really matter here (OK, a much stronger machine would be nice ...).
60b-Selfplay used KataGo 13.x with suitable CUDA / CuDNN, 28b-Selfplay uses KataGo 14.1 with suitable TensorRT.RobertJasiek wrote:Libraries: versions of CUDA, CuDNN, TensorRT.
Software: versions of KataGo and what model net, which Go GUI for calling KataGo.
The 60b/28b vs. 40b games are played with KataGo 14.1 with suitable CUDA / CuDNN via Sabaki.
+ + + + + + + + + +
I think that Karl created a new 40b-version about every day.
A new 60b-version was once created every two to three days. Now (after several hardware accidents) with a weaker machine than in the beginning a new 28b-version is created every week.
Starting points were the strongest "normal" 60b / 28b nets available at that time in the KataGo repository.
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
This diagram shows the development of the number of mistakes when following the favoured moves in Sabaki's ANALYSIS mode (starting from the problem diagram) over the version / generation of KataGo's net.
Blue is 60b, red is 28b.
+ + + + + + + + + +
In the non-Main-Semeai variations posted above, a state has been reached in which there are only one or two lost games of 28b in the 16 28b/40b games per generation of the net (or none of course).
Typical remaining mistakes of KataGo during PLAY with this problem are
Blue is 60b, red is 28b.
+ + + + + + + + + +
In the non-Main-Semeai variations posted above, a state has been reached in which there are only one or two lost games of 28b in the 16 28b/40b games per generation of the net (or none of course).
Typical remaining mistakes of KataGo during PLAY with this problem are
- initiating a fake CHITOOM (not always playing the very same moves is not always the best policy here
) - playing a PASS too early
- miscalculating the relationship of Main-Semeai-Liberties when on unfamiliar ground
- playing a fake forcing move
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
084_E15
The Guzumi
has been played just before the termination of the Crosscut Sequence in the upper left centre.
Some time earlier, White answered Black's push
into her upper right corner with
. The rest of that reducition sequence
/
also took place here.
Finally, White pushed into Black's upper left corner from below with
.
Please note that the Guzumi could also be played one move earlier. In this case, the forcing exchange of
/
would be played immediately after
/
, not before.
The diagrams below show the most played lines in the games between the 28b and the 40b net.
Black = 28b / White = 40b Black = 40b / White = 28b
Those who want to study this non-Main-Semeai variation any further on their own will find below compilations of all 28b/40b games ending in Jigo.
The comments indicate the number of games in which the respective variation was selected.
E.g. "A=17 // B=23" == Variation "A" was played in 17 games, variation "B" in 23 games.
(The labelling is designed for SGF-editors that label variations with capitals "A", "B", "C", ... Here, in the forum's SGF-browser's diagrams, you will have to take "1" as "A", "2" as "B, "3" as "C", ...)
The Guzumi
Some time earlier, White answered Black's push
Finally, White pushed into Black's upper left corner from below with
Please note that the Guzumi could also be played one move earlier. In this case, the forcing exchange of
The diagrams below show the most played lines in the games between the 28b and the 40b net.
Black = 28b / White = 40b Black = 40b / White = 28b
Those who want to study this non-Main-Semeai variation any further on their own will find below compilations of all 28b/40b games ending in Jigo.
The comments indicate the number of games in which the respective variation was selected.
E.g. "A=17 // B=23" == Variation "A" was played in 17 games, variation "B" in 23 games.
(The labelling is designed for SGF-editors that label variations with capitals "A", "B", "C", ... Here, in the forum's SGF-browser's diagrams, you will have to take "1" as "A", "2" as "B, "3" as "C", ...)
- Attachments
-
- IgoHatsuyoron120_084_E15_28bWhite#199.sgf
- (17.36 KiB) Downloaded 295 times
-
- IgoHatsuyoron120_084_E15_28bBlack#199.sgf
- (17.76 KiB) Downloaded 304 times
Last edited by Cassandra on Sun Feb 09, 2025 10:29 pm, edited 1 time in total.
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
082_E15
The Guzumi
has been played just before the termination of the Crosscut Sequence in the upper left centre.
Some time earlier, White answered Black's push
into her upper right corner with the "common" block
. The rest of that reducition sequence
/
also took place here.
Finally, White pushed into Black's upper left corner from below with
.
Please note that the Guzumi could also be played one move earlier. In this case, the forcing exchange of
/
would be played immediately after
/
, not before.
The diagrams below show the most played lines in the games between the 28b and the 40b net.
Black = 28b / White = 40b Black = 40b / White = 28b
Those who want to study this non-Main-Semeai variation any further on their own will find below compilations of all 28b/40b games ending in Jigo.
The comments indicate the number of games in which the respective variation was selected.
E.g. "A=13 // B=12 // C=279" == Variation "A" was played in 13 games, variation "B" in 12 games, and variation "C" in 279 games.
(The labelling is designed for SGF-editors that label variations with capitals "A", "B", "C", ... Here, in the forum's SGF-browser's diagrams, you will have to take "1" as "A", "2" as "B, "3" as "C", ...)
The Guzumi
Some time earlier, White answered Black's push
Finally, White pushed into Black's upper left corner from below with
Please note that the Guzumi could also be played one move earlier. In this case, the forcing exchange of
The diagrams below show the most played lines in the games between the 28b and the 40b net.
Black = 28b / White = 40b Black = 40b / White = 28b
Those who want to study this non-Main-Semeai variation any further on their own will find below compilations of all 28b/40b games ending in Jigo.
The comments indicate the number of games in which the respective variation was selected.
E.g. "A=13 // B=12 // C=279" == Variation "A" was played in 13 games, variation "B" in 12 games, and variation "C" in 279 games.
(The labelling is designed for SGF-editors that label variations with capitals "A", "B", "C", ... Here, in the forum's SGF-browser's diagrams, you will have to take "1" as "A", "2" as "B, "3" as "C", ...)
- Attachments
-
- IgoHatsuyoron120_082_E15_28bWhite#203.sgf
- (111.28 KiB) Downloaded 274 times
-
- IgoHatsuyoron120_082_E15_28bBlack#203.sgf
- (128.02 KiB) Downloaded 285 times
Last edited by Cassandra on Sun Feb 09, 2025 10:31 pm, edited 1 time in total.
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
082_E15
Both side's option for returning to the Main Semeai
In this non-Main-Semeai Variation, both sides may cooperate to return to the Main Semeai.
White connects her single stone in the upper right centre with
, instead of following the non-Main-Semeai path with
.
Black blocks with
in the upper left corner, instead of following the non-Main-Semeai path by securing two eyes for his large group with a move at a at the right edge.
...
Both side's option for returning to the Main Semeai
In this non-Main-Semeai Variation, both sides may cooperate to return to the Main Semeai.
White connects her single stone in the upper right centre with
Black blocks with
...
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: The most difficult problem ever. Igo Hatsuyoron 120
The other non-Main-Semeai variations
The other non-Main-Semeai variations still have a long way to go before consolidation, as the board is much emptier / more open and there are therefore many more variations to investigate.
046048_F2J3
Really seems to come down to Jigo.
However, the question of whether
is actually the best answer to
still needs to be clarified.
046_J3
Seems to be a bit easier for Black than for White.
046_F2
Easier for Black than for White.
A move at a seems to be an interesting option for Black now.
044_K3
Easier for Black than for White.
040_K3
Easier for White than for Black.
036_L3
Seems to come down to Jigo.
036_F2
Probably a "false positive".
BOTH nets win a lot of games with Black.
012_O5
Easier for Black than for White.
011_O8
Seems to be a bit easier for Black than for White.
010_O5
Seems to be a bit easier for Black than for White.
009_N9
Probably a "false positive".
BOTH nets win a lot of games with Black.
008_O5
Seems to be a bit easier for Black than for White.
The other non-Main-Semeai variations still have a long way to go before consolidation, as the board is much emptier / more open and there are therefore many more variations to investigate.
046048_F2J3
Really seems to come down to Jigo.
However, the question of whether
046_J3
Seems to be a bit easier for Black than for White.
046_F2
Easier for Black than for White.
A move at a seems to be an interesting option for Black now.
044_K3
Easier for Black than for White.
040_K3
Easier for White than for Black.
036_L3
Seems to come down to Jigo.
036_F2
Probably a "false positive".
BOTH nets win a lot of games with Black.
012_O5
Easier for Black than for White.
011_O8
Seems to be a bit easier for Black than for White.
010_O5
Seems to be a bit easier for Black than for White.
009_N9
Probably a "false positive".
BOTH nets win a lot of games with Black.
008_O5
Seems to be a bit easier for Black than for White.
Last edited by Cassandra on Mon Feb 10, 2025 8:13 am, edited 1 time in total.
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
-
RobertJasiek
- Judan
- Posts: 6273
- Joined: Tue Apr 27, 2010 8:54 pm
- GD Posts: 0
- Been thanked: 797 times
- Contact:
Re: The most difficult problem ever. Igo Hatsuyoron 120
Sorry for playing the noob.
Why would 36 at 41 fail?
Why does White postpone dissolution of the hane-seki as long as possible?
If White will dissolve the hane-seki, why does he play all the approach moves in the upper right?
Why would 36 at 41 fail?
Why does White postpone dissolution of the hane-seki as long as possible?
If White will dissolve the hane-seki, why does he play all the approach moves in the upper right?