More Shibano AI magic
-
Uberdude
- Judan
- Posts: 6727
- Joined: Thu Nov 24, 2011 11:35 am
- Rank: UK 4 dan
- GD Posts: 0
- KGS: Uberdude 4d
- OGS: Uberdude 7d
- Location: Cambridge, UK
- Has thanked: 436 times
- Been thanked: 3718 times
Re: More Shibano AI magic
Bill, of the various bots around, I think Elf is more side-oriented than most, in that other bots like LZ will dislike the side extension more (in relative terms, maybe not in percentage points as Elf has big percentage swings). I noticed that Elf quite likes the 3 space high pincer to 4-4 approaches whereas LZ did not ~1+ year ago. LZ has become more partial to them, I suspect in part due to some Elf games in the training, and in part naturally discovering that it is sometimes a good move and moreso than it previously thought.
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: More Shibano AI magic
Thanks much.Uberdude wrote:Bill, of the various bots around, I think Elf is more side-oriented than most, in that other bots like LZ will dislike the side extension more (in relative terms, maybe not in percentage points as Elf has big percentage swings). I noticed that Elf quite likes the 3 space high pincer to 4-4 approaches whereas LZ did not ~1+ year ago. LZ has become more partial to them, I suspect in part due to some Elf games in the training, and in part naturally discovering that it is sometimes a good move and moreso than it previously thought.
As Keynes pointed out in his Treatise on Probability many years ago, generalization requires differences in conditions, which, OC, was hardly a new thought at the time. 
A major reason for my Opening problems for AI series is to get differing opinions of different bots. I think that they are more likely to agree on human mistakes than on which plays are best. As for Elf liking the sides more than other bots, that's interesting. Elf regards some of Takagawa's side extensions as losing more than 10%.
What must LZ think?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.
-
Harleqin
- Lives in sente
- Posts: 921
- Joined: Sat Mar 06, 2010 10:31 am
- Rank: German 2 dan
- GD Posts: 0
- Has thanked: 401 times
- Been thanked: 164 times
Re: More Shibano AI magic
It may be more of a tangent, or a continuation to the next move, but how does Go Seigen's “White 8 is bad” fit in here?
A good system naturally covers all corner cases without further effort.
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: More Shibano AI magic
Or maybe thisHarleqin wrote:It may be more of a tangent, or a continuation to the next move, but how does Go Seigen's “White 8 is bad” fit in here?
?In Go Seigen: 21st Century Go, vol. 10, p. 126, Go Seigen says that he does not care for it, and gives a long sequence starting with Black b that he likes for Black. For
Go Seigen prefers a. Elf also prefers a, but only by ½%. Maybe other bots have a stronger preference.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.
-
John Fairbairn
- Oza
- Posts: 3724
- Joined: Wed Apr 21, 2010 3:09 am
- Has thanked: 20 times
- Been thanked: 4672 times
Re: More Shibano AI magic
Many a true word is spoken in jest. I have done a lot to propagate the joke, but it was T Mark Hall who first made the observation. Over the years we discovered many examples, and in fact one of the highlights of our weekly meetings in a Korean restaurant was sharing the latest sightings.It may be more of a tangent, or a continuation to the next move, but how does Go Seigen's “White 8 is bad” fit in here?
With Mark's death in 2013, the hunt came to an end. But in 2016 (just before AI rocked the boat) Ishida Yoshio published a book that wasn't a million miles from the same idea. It was a book about how changing the kakari changes the game and (I'm quoting from memory here) the most decisive point in the game was therefore move 5. His examples ranged beyond move 5, as it happens, but it was always a move very, very early in the game.
Again from memory, Ishida wasn't saying one move was better than an another. He was rather trying to show the reader how to think about the opening. The fundamental point, I think, was that a move very early in the game had a proportionally higher effect than a move later on. Thus, on move 5 (ratio 3:2) the effect of the extra Black stone was much greater than on, say, move 25 (13:12). There were, however, other books around the same time that played the same sort of numbers game even quite late in the fuseki. I think it was Mizokami who demonstrated a method of choosing how to play (aggressively, neutrally or cautiously) by counting the proportions of opposing stones in various areas of the board.
In Ishida's case, however, having a large effect so early in the game is tantamount to having control of the game. It may be, therefore, that we are seeing something similar in AI play. The Direct 3-3 on move 5 may not be inherently any better than the same move on move 25, but what is better is simply having some degree of control of the game. I hasten to add that control doesn't have to mean bossing the opponent round. It may, for example, mean making probes - if you've got better maps, you can get control of the high ground later on.
I have a hunch that a lot of AI play can be explained in such simplistic terms. After all, as Michael Redmond wisely said, go is really a rather simple game. It's rather like when the military strategicians turn to Sun Zi's Art of War. In Sun Zi's time, warfare was relatively simple - no cyber this, no bio that. No numbers! At any rate, it's easier to spot controlling factors with a simplified layout.
Going back to why White 8 was so often bad in reality as in jest, I think it might be fair to say it was because White was trying to wrest control of the game far too son: when the proportions were still 4:4 followed by 5:4, as opposed 20:20 followed by 21:20.
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: More Shibano AI magic
I seem to have lost a note. Did I accidentally delete it?
Let's try again.
In Go Seigen: 21st Century Go, vol. 4, p. 18, Go Seigen says that a lot of people play this 3-3 invasion to the one space pincer, but that it is not good. He gives this diagram. He goes on to say, if you like the 3-3 invasion so well, why not just play it directly? Thus anticipating the bots once again.
Elf agrees that this diagram is not good for White, but thinks that the problem is
, which loses 8% to the sagari in the next diagram.
Black gets a nice wall, but has an open skirt.
Let's try again.
In Go Seigen: 21st Century Go, vol. 4, p. 18, Go Seigen says that a lot of people play this 3-3 invasion to the one space pincer, but that it is not good. He gives this diagram. He goes on to say, if you like the 3-3 invasion so well, why not just play it directly? Thus anticipating the bots once again.
Elf agrees that this diagram is not good for White, but thinks that the problem is
, which loses 8% to the sagari in the next diagram.Black gets a nice wall, but has an open skirt.
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.
-
sorin
- Lives in gote
- Posts: 389
- Joined: Wed Apr 21, 2010 9:14 pm
- Has thanked: 418 times
- Been thanked: 198 times
Re: More Shibano AI magic
I think that the answer to that question has to do with the blind-spot about the 3-3 invasion: before AlphaGo, everyone thought that a 3-3 move in response to 4-4 is a passive invasion, while AI showed us that it is in fact an active move (almost an attacking move ?).Bill Spight wrote:Maybe we have it backwards. Maybe the question is why we came to believe that the second move in the corner is unnecessary.Knotwilg wrote:That has a circular scent. Isn't the keima approach is a desirable move because it prevents the owner of the corner stone to make a very desirable move, the enclosure?sorin wrote: The commonality I see between the 3-4 based shimari and the 4-4 based one is that they both prevent the opponent from playing a very desirable move (a keima approach).
As Uberdude has pointed out in his "LeelaZero opening gospel", AI thinks corners are urgent, not only the first move in the corner but also the second move. Now the question is why?
Sorin - 361points.com
-
Kirby
- Honinbo
- Posts: 9553
- Joined: Wed Feb 24, 2010 6:04 pm
- GD Posts: 0
- KGS: Kirby
- Tygem: 커비라고해
- Has thanked: 1583 times
- Been thanked: 1707 times
Re: More Shibano AI magic
My version of KataGo agrees with your last diagram. However, what I also find surprising is that it seems to prefer black's block on *this* side:
I stopped after about 12k playouts. The variation it gives is this one:
I stopped after about 12k playouts. The variation it gives is this one:
be immersed
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: More Shibano AI magic
Thanks, Kirby. I have noticed that pattern, as well. But just because Elf agrees with itself means little.
The first strange thing is that Black seems to block on the wrong side. We're already used to the block on the unexpected side against the direct 3-3, but surely Black wants to make a wall with the wide extension.
Part of the puzzle is resolved by the fact that Black does not protect against the cut. Black leaves the turn as an option. But, OC, Black can tenuki without playing the turn.
It's a New World, isn't it? 
The first strange thing is that Black seems to block on the wrong side. We're already used to the block on the unexpected side against the direct 3-3, but surely Black wants to make a wall with the wide extension.
Part of the puzzle is resolved by the fact that Black does not protect against the cut. Black leaves the turn as an option. But, OC, Black can tenuki without playing the turn. It's a New World, isn't it?
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.
-
Kirby
- Honinbo
- Posts: 9553
- Joined: Wed Feb 24, 2010 6:04 pm
- GD Posts: 0
- KGS: Kirby
- Tygem: 커비라고해
- Has thanked: 1583 times
- Been thanked: 1707 times
Re: More Shibano AI magic
It is, yeah. I'm confused, but I think I'm starting to enjoy it.Bill Spight wrote: It's a New World, isn't it?
be immersed
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: More Shibano AI magic
I think I have a better example for what I was trying to say about Shibano's argument.
I am reasonably certain how LZ would play, following the Gospel.
First the press, then the enclosure.
Well, kowabunga, that's how Elf would play, as well. And my guess is so would KataGo and other top bots.
But Cho Chikun was White, vs. Rin Kaiho (GoGoD 1994-03-03a), and he made a textbook play, if not the textbook play.
After all, Black's formation on the left side cries out for a Black stone on K-04 to make a double wing and a huge moyo, and
prevents that, besides making an extension for White. The direction of play dictates such an extension.
Well, the bots are teaching us different. And I am quite willing to concede that they are very probably right, and the Gospel sequence dominates the bottom extension. But to dominate the extension, the Gospel sequence only has to be at least as good. It does not have to be a lot better, and it does not have to have a winrate difference greater than the margin of error. (OC, if we only had one such example, we would want that.) Elf calculates a winrate difference of only 2%, surely well within its margin of error. Since Elf tends to find larger winrate differences than other top bots, I would be surprised if either LZ or KataGo evaluates the extension as more than 2% worse than the press or the enclosure.
My point is this. To find that the press or the enclosure is superior to the extension, we only have to show how good they are, not how bad the extension is. Particularly as, even in this case, which seems clear, the extension is very likely to be playable. The press and the extension are better than we thought. We don't have to find reasons to say that the extension is worse than we thought. Maybe it isn't.
I am reasonably certain how LZ would play, following the Gospel.
First the press, then the enclosure.
Well, kowabunga, that's how Elf would play, as well. And my guess is so would KataGo and other top bots.
But Cho Chikun was White, vs. Rin Kaiho (GoGoD 1994-03-03a), and he made a textbook play, if not the textbook play.
After all, Black's formation on the left side cries out for a Black stone on K-04 to make a double wing and a huge moyo, and
prevents that, besides making an extension for White. The direction of play dictates such an extension.Well, the bots are teaching us different. And I am quite willing to concede that they are very probably right, and the Gospel sequence dominates the bottom extension. But to dominate the extension, the Gospel sequence only has to be at least as good. It does not have to be a lot better, and it does not have to have a winrate difference greater than the margin of error. (OC, if we only had one such example, we would want that.
) Elf calculates a winrate difference of only 2%, surely well within its margin of error. Since Elf tends to find larger winrate differences than other top bots, I would be surprised if either LZ or KataGo evaluates the extension as more than 2% worse than the press or the enclosure. My point is this. To find that the press or the enclosure is superior to the extension, we only have to show how good they are, not how bad the extension is. Particularly as, even in this case, which seems clear, the extension is very likely to be playable. The press and the extension are better than we thought. We don't have to find reasons to say that the extension is worse than we thought. Maybe it isn't.
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.
-
Kirby
- Honinbo
- Posts: 9553
- Joined: Wed Feb 24, 2010 6:04 pm
- GD Posts: 0
- KGS: Kirby
- Tygem: 커비라고해
- Has thanked: 1583 times
- Been thanked: 1707 times
Re: More Shibano AI magic
I looked at this position with KataGo a little bit as well.
It was a little bit interesting when I adjusted the komi. When I have komi as 7.5, KataGo thinks white is winning, and recommends the press you mentioned, without considering many other alternatives: Above, it thinks white has about a 57% chance of winning, and seems comfortable recommending the press.
But if I lower the komi to 6.5, it thinks white has only about a 47% chance of winning, and now starts thinking of invasion: It still kind of likes the press, but there's more "thought" given to this invasion on the left. When I reduce the komi to something like 2.5, it likes the invasion on the left even more. That being said, it really thinks white is losing in such a position. But still interesting to me.
It was a little bit interesting when I adjusted the komi. When I have komi as 7.5, KataGo thinks white is winning, and recommends the press you mentioned, without considering many other alternatives: Above, it thinks white has about a 57% chance of winning, and seems comfortable recommending the press.
But if I lower the komi to 6.5, it thinks white has only about a 47% chance of winning, and now starts thinking of invasion: It still kind of likes the press, but there's more "thought" given to this invasion on the left. When I reduce the komi to something like 2.5, it likes the invasion on the left even more. That being said, it really thinks white is losing in such a position. But still interesting to me.
be immersed
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: More Shibano AI magic
Yeah, me too. Many thanks.Kirby wrote:I looked at this position with KataGo a little bit as well.
It was a little bit interesting when I adjusted the komi. When I have komi as 7.5, KataGo thinks white is winning, and recommends the press you mentioned, without considering many other alternatives:
{snip}
Above, it thinks white has about a 57% chance of winning, and seems comfortable recommending the press.
But if I lower the komi to 6.5, it thinks white has only about a 47% chance of winning, and now starts thinking of invasion:
{snip}
When I reduce the komi to something like 2.5, it likes the invasion on the left even more. That being said, it really thinks white is losing in such a position. But still interesting to me.
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.
-
Knotwilg
- Oza
- Posts: 2432
- Joined: Fri Jan 14, 2011 6:53 am
- Rank: KGS 2d OGS 1d Fox 4d
- GD Posts: 0
- KGS: Artevelde
- OGS: Knotwilg
- Online playing schedule: UTC 18:00 - 22:00
- Location: Ghent, Belgium
- Has thanked: 360 times
- Been thanked: 1021 times
- Contact: