alreadydone,
1) thank you for your links in your message body.
2) Your two attachments contain photo/scan copies of several pages of a book. IMO, the copies have more pages than can be expected to be fair use when spread publicly. Judging from experience here, I'd say, one or two pages of photo/scan copies would be ok. Note that citations (that is not: photo/scan copies) with proper credit would allow more pages if occurring in new printed media (that is not: wiki web pages).
KataGo @ Igo Hatsuyoron 120
-
RobertJasiek
- Judan
- Posts: 6272
- Joined: Tue Apr 27, 2010 8:54 pm
- GD Posts: 0
- Been thanked: 797 times
- Contact:
-
hyperpape
- Tengen
- Posts: 4382
- Joined: Thu May 06, 2010 3:24 pm
- Rank: AGA 3k
- GD Posts: 65
- OGS: Hyperpape 4k
- Location: Caldas da Rainha, Portugal
- Has thanked: 499 times
- Been thanked: 727 times
Re: KataGo @ Igo Hatsuyoron 120
Did you by any chance test the version trained on Igo Hatsuyoron against any other versions of Kata? I don't know that it's an important question, but it piqued my curiosity whether it learned anything generalizable, forgot something other versions knew, or was essentially unaffected by spending so much time on such an obscure scenario.
-
dany
- Dies with sente
- Posts: 121
- Joined: Fri Dec 20, 2019 2:44 am
- Rank: IGS 4 dan
- GD Posts: 0
- Has thanked: 36 times
- Been thanked: 12 times
Re: KataGo @ Igo Hatsuyoron 120
- Attachments
-
- igo120ko_var4.sgf
- (1.93 KiB) Downloaded 646 times
-
- igo120ko_var3.sgf
- (2.09 KiB) Downloaded 646 times
-
- igo120ko_var1.sgf
- (2.08 KiB) Downloaded 636 times
-
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: KataGo @ Igo Hatsuyoron 120
Dear dany,
Thank you very much indeed for your very interesting suggestion.
To be honest, I have not checked your ko variations, because usually Black is unable to win any ko in this problem.
In my opinion (not yet crosschecked by Joachim), Black can (should) do without your ko, and still reaches a jigo (with your choice of setting the komi to zero).
To ease the evaluation of "your" position (and the variations thereafter), I have played all the other "unaffected" moves of the Main Semeai first, and reached the following state on the board:
Please be aware:
Main points of interest are:
Thank you very much indeed for your very interesting suggestion.
To be honest, I have not checked your ko variations, because usually Black is unable to win any ko in this problem.
In my opinion (not yet crosschecked by Joachim), Black can (should) do without your ko, and still reaches a jigo (with your choice of setting the komi to zero).
To ease the evaluation of "your" position (and the variations thereafter), I have played all the other "unaffected" moves of the Main Semeai first, and reached the following state on the board:
Please be aware:
- White will never capture with moves at any of the b-points at the left.
- However, Black will have to connect at these points, if necessary, before White's main decision point (i.e. just before she captures the hanezeki's tail, or chooses the Semeai Variation).
- Black will never play at any of the w-points at the right, after White's decision between the Capture and the Semeai Variation.
- White will capture the hanezeki's tail, or choose to continue with the Semeai Variation, at the moment when Black's large group at the top right has only seven liberties left -- after a Black move.
Main points of interest are:
- Will White connect at a at the upper edge, or will Black capture White's single stone there?
- Will White play the kikashi of W k & W c in the lower left corner, or will Black be able to defend at c?
- Attachments
-
- danysKo.sgf
- (29.09 KiB) Downloaded 1437 times
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: KataGo @ Igo Hatsuyoron 120
Dany kindly made me aware that I missed an important detail ...
After dany's creation of the ko shape in the centre, Black is unable to strive for a jigo (komi = 0) by definding his lower left corner with the move at 4.
Due to the shortage of liberties of Black's two stones at the left, Black is forced to start a ko fight with 8 that he cannot win.
With what we know now, the variation with Black's tenuki after KataGo's bad-shape move really ends with W + 1 (komi = 0) respectively W + 2 (komi = 1).
After dany's creation of the ko shape in the centre, Black is unable to strive for a jigo (komi = 0) by definding his lower left corner with the move at 4.
Due to the shortage of liberties of Black's two stones at the left, Black is forced to start a ko fight with 8 that he cannot win.
With what we know now, the variation with Black's tenuki after KataGo's bad-shape move really ends with W + 1 (komi = 0) respectively W + 2 (komi = 1).
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: KataGo @ Igo Hatsuyoron 120
Three weeks ago, my laptop said "Good bye!" to me. A sustainable repair was no longer possible, but luckily it could be restored at least to the extent that I could read out the data on the hard disk and save it. Otherwise, about six weeks of analytic and preparatory work for our website since the last backup would have been lost.
I had a lot of time to think during these three weeks, and now could be a good time to share our findings and results so far.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
In January, dany kindly introduced his atekomi W 1, resulting in a ko fight in the left centre after B 2 amd W 3.
This ko fight will end with White winning both Main Variations (Capture & Semeai) by two points (= W+2 / W+2 @ komi=1).
Since then, we were busy with investigating the question whether Black might be better in the ko fight, if he had not played his kikashi (shadowed board points) beforehand, and could use these as ko threats.
We used this position as a general starting point for our "no kikashi" variation tree.
To keep a long story short:
Black cannot benefit from keeping his kikashi in reserve, but he will give White chances, and increase the complexity of the variation tree dramatically.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
You are probably familiar with the wording of the Pareto Principle, according to which the last 20 percent of the work takes 80 percent of the time.
I have experienced several of these 80/20-moments, when I thought that the end would be near, but when emails from dany, or Joachim, arrived that questioned much of the work so far.
Additionally, you won't get dumber while painting, and commenting, diagrams, so the number of errors found increased, as did the number of variations.
In retrospect, many of our (mainly mine) initial mistakes might look a bit silly, but this cannot be helped.
I had a lot of time to think during these three weeks, and now could be a good time to share our findings and results so far.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
In January, dany kindly introduced his atekomi W 1, resulting in a ko fight in the left centre after B 2 amd W 3.
This ko fight will end with White winning both Main Variations (Capture & Semeai) by two points (= W+2 / W+2 @ komi=1).
Since then, we were busy with investigating the question whether Black might be better in the ko fight, if he had not played his kikashi (shadowed board points) beforehand, and could use these as ko threats.
We used this position as a general starting point for our "no kikashi" variation tree.
To keep a long story short:
Black cannot benefit from keeping his kikashi in reserve, but he will give White chances, and increase the complexity of the variation tree dramatically.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
You are probably familiar with the wording of the Pareto Principle, according to which the last 20 percent of the work takes 80 percent of the time.
I have experienced several of these 80/20-moments, when I thought that the end would be near, but when emails from dany, or Joachim, arrived that questioned much of the work so far.
Additionally, you won't get dumber while painting, and commenting, diagrams, so the number of errors found increased, as did the number of variations.
In retrospect, many of our (mainly mine) initial mistakes might look a bit silly, but this cannot be helped.
Last edited by Cassandra on Mon Jul 20, 2020 1:55 am, edited 3 times 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: KataGo @ Igo Hatsuyoron 120
My biggest mistake was to take the kikashi for granted.
Probably my work on the many variations was too mechanical in the very beginning.
Even with Black's shadowed kikashi unplayed, White will win the ko in the left centre.
However, in the very end, she will have to play the disadvatageous ko threat of W 1 in the lower left corner that costs her one point overall. The final scores will be W+1 / W+1 (@ komi=1).
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Dany's first suggestion was ...
... to keep White's atari during the creation of the ko shape in the left centre (W a, B b) in reserve for the later use as an additional ko threat.
Now, after White won the ko in the left centre, White is able to play the forcing move W 1 in the lower left corner later, threatening to reduce Black's large group at the left to only one eye. Black is forced to answer below.
Black does not have any eyes in the left centre, and so will have to capture White's two stones in the lower left corner with B a before the game ends.
The final scores will become W+2 / W+2 (@ komi=1) again.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Dany's second suggestion was related to a White's gote move at 1 in the hanezeki, stopping Black from playing his kikashi there.
After Black's atari 2 in the lower left corner, White might be able to sacrifice her two stones there.
White's kikashi 3 in the hanezeki gains her an additional three points in the Semeai Variation. B 4 will become an additional prisoner, while the point of a will remain White territory.
(White will lose one point in the Capture Variation by this action, but she prioritises the Semeai Variation.)
Capturing two White stones with B 6 gains two points in the Capture Variation only, but which might not help if White finally chooses the other Main Variation.
Additionally, Black does not gain two additional liberties in the Main Semeai, as with a White connection at 6. This might restrict his actions in other parts of the board.
Probably my work on the many variations was too mechanical in the very beginning.
Even with Black's shadowed kikashi unplayed, White will win the ko in the left centre.
However, in the very end, she will have to play the disadvatageous ko threat of W 1 in the lower left corner that costs her one point overall. The final scores will be W+1 / W+1 (@ komi=1).
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Dany's first suggestion was ...
... to keep White's atari during the creation of the ko shape in the left centre (W a, B b) in reserve for the later use as an additional ko threat.
Now, after White won the ko in the left centre, White is able to play the forcing move W 1 in the lower left corner later, threatening to reduce Black's large group at the left to only one eye. Black is forced to answer below.
Black does not have any eyes in the left centre, and so will have to capture White's two stones in the lower left corner with B a before the game ends.
The final scores will become W+2 / W+2 (@ komi=1) again.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Dany's second suggestion was related to a White's gote move at 1 in the hanezeki, stopping Black from playing his kikashi there.
After Black's atari 2 in the lower left corner, White might be able to sacrifice her two stones there.
White's kikashi 3 in the hanezeki gains her an additional three points in the Semeai Variation. B 4 will become an additional prisoner, while the point of a will remain White territory.
(White will lose one point in the Capture Variation by this action, but she prioritises the Semeai Variation.)
Capturing two White stones with B 6 gains two points in the Capture Variation only, but which might not help if White finally chooses the other Main Variation.
Additionally, Black does not gain two additional liberties in the Main Semeai, as with a White connection at 6. This might restrict his actions in other parts of the board.
Last edited by Cassandra on Mon Jul 20, 2020 3:55 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: KataGo @ Igo Hatsuyoron 120
Dany's third suggestion concerned a split of the Nakade Sequence (that is used as first White ko threat).
Dany idea was to keep the final exchange of W a, B b in reserve (e.g. for the later usage as an additional ko threat), and to recapture the ko as early as with W 4.
Indeed, if Black continued with fighting the ko -- using his kikashi in the hanezeki as a ko threat -- White would achieve her aim by also using the continuation of the Nakade Seqeunce as a ko threat. Black has fallen into a trap, as he unnecessarily wasted his ko threat at B 5 -- compared with the continuesly played Nakade Sequence (i.e. W 4 at 8).
Eventually, it was Joachim, who found the correct continuation for Black.
Dany idea was to keep the final exchange of W a, B b in reserve (e.g. for the later usage as an additional ko threat), and to recapture the ko as early as with W 4.
Indeed, if Black continued with fighting the ko -- using his kikashi in the hanezeki as a ko threat -- White would achieve her aim by also using the continuation of the Nakade Seqeunce as a ko threat. Black has fallen into a trap, as he unnecessarily wasted his ko threat at B 5 -- compared with the continuesly played Nakade Sequence (i.e. W 4 at 8).
Eventually, it was Joachim, who found the correct continuation for Black.
Last edited by Cassandra on Mon Jul 20, 2020 5:39 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: KataGo @ Igo Hatsuyoron 120
Please assume that Black was able to connect his single stone in the lower right with the triangled stone AND that Black's large group in the upper right has NINE Main Semeai Liberties at White's decision point.
If Black had TEN Main Semeai Liberties, White would be unable to capture this Black group in the Semeai Variation.
White chooses the Semeai Variation.
At the last possible moment, Black captures in the upper right.
It will not benefit White to connect at a (Black will capture at b thereafter), instead of occupying a Black liberty with 21, just because White does not have any relevant ko threats available.
White is unable to answer Black's ko threat at the upper edge, because Black would give atari at a, saving his large group in question.
Black captures with 28 at the upper edge, winning the game by about 15 points.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
This result implies that White must not play her last gote move, e.g. X.
White's move 1 occupied one of Black's liberties in the upper right, instead.
But then Black could capture at the upper edge ...
... and connect thereafter. Black gained two points overall.
If Black had TEN Main Semeai Liberties, White would be unable to capture this Black group in the Semeai Variation.
White chooses the Semeai Variation.
At the last possible moment, Black captures in the upper right.
It will not benefit White to connect at a (Black will capture at b thereafter), instead of occupying a Black liberty with 21, just because White does not have any relevant ko threats available.
White is unable to answer Black's ko threat at the upper edge, because Black would give atari at a, saving his large group in question.
Black captures with 28 at the upper edge, winning the game by about 15 points.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
This result implies that White must not play her last gote move, e.g. X.
White's move 1 occupied one of Black's liberties in the upper right, instead.
But then Black could capture at the upper edge ...
... and connect thereafter. Black gained two points overall.
Last edited by Cassandra on Mon Jul 20, 2020 8:20 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: KataGo @ Igo Hatsuyoron 120
There is a similar effect on Black's side, which is likely to be known already.
If -- at White's decision point -- Black has only SEVEN Main Semeai Liberties, the "natural" looking move of B 1, capturing White's single stone at the upper edge, is a grave mistake.
By this move, Black has fallen below the Punishment Semeai Liberty Level. He should have played oki at the left edge.
White's blocking move at 2 -- also looking "natural" -- is also a grave mistake, giving the favour back to Black.
White must continue with the Main Semeai, instead, occupying one of Black's liberties in the top right.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Everything is clearly visible at White's decision point.
But what about the period before, back until Black's move 117?
Who can immediately say with certainty whether Black's move at 1 is possible or not?
If -- at White's decision point -- Black has only SEVEN Main Semeai Liberties, the "natural" looking move of B 1, capturing White's single stone at the upper edge, is a grave mistake.
By this move, Black has fallen below the Punishment Semeai Liberty Level. He should have played oki at the left edge.
White's blocking move at 2 -- also looking "natural" -- is also a grave mistake, giving the favour back to Black.
White must continue with the Main Semeai, instead, occupying one of Black's liberties in the top right.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Everything is clearly visible at White's decision point.
But what about the period before, back until Black's move 117?
Who can immediately say with certainty whether Black's move at 1 is possible or not?
Last edited by Cassandra on Fri Jul 24, 2020 8:31 am, edited 2 times 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)
-
yoyoma
- Lives in gote
- Posts: 653
- Joined: Mon Apr 19, 2010 8:45 pm
- GD Posts: 0
- Location: Austin, Texas, USA
- Has thanked: 54 times
- Been thanked: 213 times
Re: KataGo @ Igo Hatsuyoron 120
You credited some people with pointing out ideas. What about Katago, is it helping to analyze some of these ideas?
-
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: KataGo @ Igo Hatsuyoron 120
Thank you very much for your interest, and especially for your question about KataGo.yoyoma wrote:You credited some people with pointing out ideas. What about Katago, is it helping to analyze some of these ideas?
Indeed, most of our work was done manually. I used KataGo sometimes for confirming a new line, when I was very unsure about it. Dany once mentioned that KataGo has confirmed his usage of ko threats during his ko. Joachim did not have a suitable machine for running KataGo. But let me explain in more detail ...
The usage of KataGo was severly affected by the following aspects:
- Neither Joachim nor me own that powerful type of machine, which would be necessary to achieve an adequate amount of playouts in justifiable time. Please note that it would be necessary in most cases to analyse the same positition with different komi, in order to get the "optimal" suggestion by KataGo, also with a certain relationship to the (expected) final score. Our machines were simply much too slow / unsuitable for that purpose.
- KataGo 1.2, which was used to create the special IgoHatsuyoron120-network, calculates with Area Scoring only. This implies that KataGo 1.2 is unable to distinguish between (e.g.) "W+1" and "W+2" under Territory Scoring. But which would be necessary for a "really" reliable analysis of a problem that originated in Japan.
- As far as I know, KataGo does not use the (estimated) final score of a variation for prioritisation. Which is our parameter for ranking subvariations.
- Additionally, (I suppose that) KataGo does not use the final score of BOTH Main Variations as parameters for ranking moves / subvariations. ONE "sure" way to (e.g.) a win for White by two points will be enough. However, we would like to prioritise "W+2 / B+4" over "W+2 / W+2" at a branching point for Black's next move.
- A technical environment that enables the creation of a reliable selection of candidate moves by KataGo in seconds, not in minutes.
- A (new?) special IgoHatsuyoron120-network that would be compatible with the newer versions of KataGo, which are able to also calculate using Territory Scoring.
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)
-
dany
- Dies with sente
- Posts: 121
- Joined: Fri Dec 20, 2019 2:44 am
- Rank: IGS 4 dan
- GD Posts: 0
- Has thanked: 36 times
- Been thanked: 12 times
Re: KataGo @ Igo Hatsuyoron 120
KataGo is a very useful tool and greatly helps in analysis. I usually find the moves manually and then check them with KataGo. This is a good way to find blunders. But KataGo's understanding of the problem is imperfect. Sometimes, for some simple positions, KataGo needs hundreds of thousands of playouts to find the right move. And there are positions where so many playouts are not enough for KataGo. Usually the best move in position is among the top 5 KataGo moves, but often it is not the top move.yoyoma wrote:What about Katago, is it helping to analyze some of these ideas?
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: KataGo @ Igo Hatsuyoron 120
I look forward to lightvector's response.Cassandra wrote:Thank you very much for your interest, and especially for your question about KataGo.yoyoma wrote:You credited some people with pointing out ideas. What about Katago, is it helping to analyze some of these ideas?
Indeed, most of our work was done manually. I used KataGo sometimes for confirming a new line, when I was very unsure about it. Dany once mentioned that KataGo has confirmed his usage of ko threats during his ko. Joachim did not have a suitable machine for running KataGo. But let me explain in more detail ...
The usage of KataGo was severly affected by the following aspects:If one wanted to have an assistant for creating these "Why not play there?" moments (which I encountered during painting the diagrams, and which seems to be one of Joachim's talent when refuting lines of play in question), I think that the following requirements were needed:
- Neither Joachim nor me own that powerful type of machine, which would be necessary to achieve an adequate amount of playouts in justifiable time. Please note that it would be necessary in most cases to analyse the same positition with different komi, in order to get the "optimal" suggestion by KataGo, also with a certain relationship to the (expected) final score. Our machines were simply much too slow / unsuitable for that purpose.
- KataGo 1.2, which was used to create the special IgoHatsuyoron120-network, calculates with Area Scoring only. This implies that KataGo 1.2 is unable to distinguish between (e.g.) "W+1" and "W+2" under Territory Scoring. But which would be necessary for a "really" reliable analysis of a problem that originated in Japan.
- As far as I know, KataGo does not use the (estimated) final score of a variation for prioritisation. Which is our parameter for ranking subvariations.
- Additionally, (I suppose that) KataGo does not use the final score of BOTH Main Variations as parameters for ranking moves / subvariations. ONE "sure" way to (e.g.) a win for White by two points will be enough. However, we would like to prioritise "W+2 / B+4" over "W+2 / W+2" at a branching point for Black's next move.
- A technical environment that enables the creation of a reliable selection of candidate moves by KataGo in seconds, not in minutes.
- A (new?) special IgoHatsuyoron120-network that would be compatible with the newer versions of KataGo, which are able to also calculate using Territory Scoring.
However, my 2¢, FWIW.Today's, top bots are not trained as analysts. Instead, they are trained to win games. The neural networks learn generally good plays and evaluations, but require tree search to learn more about specific positions. It would be good, therefore, to have a bot that has been trained on this particular position. Fortunately, lightvector supplied one.
How to use it? Given the difficulty of this problem, and the possibility, given sekis and kos, that correct play may differ between Japanese and area scoring, it does not appear that you can just let the specialized bot run for a day or two. Weeks or months may be required and then the variations would have to be checked by hand. It seems that interactive use is required, with an emphasis, to accomodate human understanding, for mostly depth first search. Also, I would try to find a optimal play by territory scoring first. That will normally be optimal by area scoring. And if not, you will know where to look for differences. Going back and forth between the two before finding optimal play by either one just makes for confusion, IMO.
As I said, I look forward to lightvector's thoughts.
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
-
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: KataGo @ Igo Hatsuyoron 120
Knowing the current status of Black's Main Semeai Liberties at every moment of the game is mandatory to identify "natural looking" moves as mistakes.
I thought that this knowledge would be helpful, at least at amateur (Kyu) level. Therefore I decided to add this "current status" to every diagram of the "no kikashi" variation tree, and to divide diagrams when this status changes. I used the following classification (measured at the start of the diagram):
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Let us accompany this marking on the way to Dany's ko.
In the starting position of our "no kikashi" analysis, Black has EIGHT Main Semeai Liberties.
This means that Black's large group in the upper right would have EIGHT liberties left at White's decision point, if -- from now on -- both sides played only moves that belong to the Main Semeai.
B 117 loses TWO Main Semeai Liberties, in priciple.
However, the marked point will become an approach-move liberty of Black's group (Black got an eye), so the effective loss in only ONE liberty.
W 118 and 120 are played outside the Main Semeai area, so Black can do also with B 119 and 121.
White's forcing exchange of W 122 for B 123 outside the Main Semeai area has been inserted to simplify the presentation.
W 124 is played outside the Main Semeai area, in principle. However, this move creates an approach-move liberty at the marked point for her group at the left side, and so does not lose any tempo in the Main Semeai.
And now, at this special moment -- what a surprise -- there will be NO ko in the left centre.
Black is unable to immediately give atari at a. This move is located outside the Main Semeai area, and would push Black below the Punishment Semeai Liberty Level.
Black is bound to choose among his moves in the Main Semeai, and giving atari at White's two stones in the lower left with B 125 is best, because White threatens to cut at b.
White's connection gives TWO Main Semeai Liberties back to Black.
Thereafter, Black has sufficient Main Semeai Liberties to start Dany's ko with B 127 and 129.
However, Black has lost one of his potential ko threats before the start of the ko fight, and so his chances to win this ko (with the marked exchange assumed to be played later).
The final scores of the game will become W+2 / W+2 (@ komi=1) again.
I thought that this knowledge would be helpful, at least at amateur (Kyu) level. Therefore I decided to add this "current status" to every diagram of the "no kikashi" variation tree, and to divide diagrams when this status changes. I used the following classification (measured at the start of the diagram):
- SIX ==> Black is below the Punishment Semeai Liberty Level, White should continue with the Main Semeai.
- SEVEN ==> Black must not play any move outside the Main Semeai area, unless White did so immediately before.
- EIGHT ==> Both sides may play with ease.
- NINE ==> White must be careful to not play any gote move in special circumstances that would allow Joachim's final ko in the Semeai Variation, if White intends to choose the Semeai Variation.
- at least TEN ==> White will lose the Semeai Variation, because she is unable to capture Black's large group in the upper right.
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Let us accompany this marking on the way to Dany's ko.
In the starting position of our "no kikashi" analysis, Black has EIGHT Main Semeai Liberties.
This means that Black's large group in the upper right would have EIGHT liberties left at White's decision point, if -- from now on -- both sides played only moves that belong to the Main Semeai.
B 117 loses TWO Main Semeai Liberties, in priciple.
However, the marked point will become an approach-move liberty of Black's group (Black got an eye), so the effective loss in only ONE liberty.
W 118 and 120 are played outside the Main Semeai area, so Black can do also with B 119 and 121.
White's forcing exchange of W 122 for B 123 outside the Main Semeai area has been inserted to simplify the presentation.
W 124 is played outside the Main Semeai area, in principle. However, this move creates an approach-move liberty at the marked point for her group at the left side, and so does not lose any tempo in the Main Semeai.
And now, at this special moment -- what a surprise -- there will be NO ko in the left centre.
Black is unable to immediately give atari at a. This move is located outside the Main Semeai area, and would push Black below the Punishment Semeai Liberty Level.
Black is bound to choose among his moves in the Main Semeai, and giving atari at White's two stones in the lower left with B 125 is best, because White threatens to cut at b.
White's connection gives TWO Main Semeai Liberties back to Black.
Thereafter, Black has sufficient Main Semeai Liberties to start Dany's ko with B 127 and 129.
However, Black has lost one of his potential ko threats before the start of the ko fight, and so his chances to win this ko (with the marked exchange assumed to be played later).
The final scores of the game will become W+2 / W+2 (@ komi=1) again.
Last edited by Cassandra on Mon Jul 20, 2020 1:14 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)