Let us have a look at this joseki:
The main variation today is:
a is the common continuation for black, b is the runner up option.
a and b have similar winrates on waltheri.
After some self reviews and a second look at my database of recent pro games, I propose a reevaluation of a and b.
In 202 games black wins 44.6% times with the move at a.
In 121 games black wins 52.1% times with the move at b.
This trend is also visible in my self reviewed games.
My conclusion:
b is my new favourite candidate move in this joseki.
What is your take on this joseki?
Joseki 4-4, low approach, low extension, contact variation
-
EdLee
- Honinbo
- Posts: 8859
- Joined: Sat Apr 24, 2010 6:49 pm
- GD Posts: 312
- Location: Santa Barbara, CA
- Has thanked: 349 times
- Been thanked: 2070 times
-
sorin
- Lives in gote
- Posts: 389
- Joined: Wed Apr 21, 2010 9:14 pm
- Has thanked: 418 times
- Been thanked: 198 times
Re:
http://ps.waltheri.net/EdLee wrote:Hi Gomoto, thanks.What's waltheri ?a and b have similar winrates on waltheri.
Sorin - 361points.com
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: Joseki 4-4, low approach, low extension, contact variati
Some months ago I discovered that the published winrates on waltheri (under the "around stones" view) were inaccurate, even randomly so. It was easy to check for options with relatively few games. But recently the software changed, and it has become more difficult to check the winrates, as the site does not publish who won or lost each game. Maybe the winrates have been corrected, but maybe not.Gomoto wrote:Let us have a look at this joseki:
The main variation today is:
a is the common continuation for black, b is the runner up option.
a and b have similar winrates on waltheri.
I would trust your results over waltheri's.After some self reviews and a second look at my database of recent pro games, I propose a reevaluation of a and b.
In 202 games black wins 44.6% times with the move at a.
In 121 games black wins 52.1% times with the move at b.
This trend is also visible in my self reviewed games.
My conclusion:
b is my new favourite candidate move in this joseki.
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.
-
EdLee
- Honinbo
- Posts: 8859
- Joined: Sat Apr 24, 2010 6:49 pm
- GD Posts: 312
- Location: Santa Barbara, CA
- Has thanked: 349 times
- Been thanked: 2070 times
-
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: Joseki 4-4, low approach, low extension, contact variati
'b' usually keeps sente, black doesn't need to descend at a after white extends: if white cuts later black extends and gets the forcing move at c. The 'b' push seems to work well when black is building on a large scale on the left side, e.g. black makes an extension, often d, after white extends. Or sometimes keeps pushing (nice if it overconcentrates white on the top).Marcel Grünauer wrote: 'b' pushes from behind and loses sente because Black still needs to play 'a'. I guess you'd play 'b' if you really don't like White to push up at 'b'.
-
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: Joseki 4-4, low approach, low extension, contact variati
8% win rate difference from a handful of human games doesn't seem like much evidence to make much of a point about a single move out of hundreds that determined the result of the game.
be immersed
-
dfan
- Gosei
- Posts: 1599
- Joined: Wed Apr 21, 2010 8:49 am
- Rank: AGA 2k Fox 3d
- GD Posts: 61
- KGS: dfan
- Has thanked: 891 times
- Been thanked: 534 times
- Contact:
Re: Joseki 4-4, low approach, low extension, contact variati
I would agree more with this if we were talking about a few dozen games rather than a few hundred. I did some simulations to ask the question "If Black won 153 out of 323 games in total and we break it into groups of 202 and 121 games, what are the chances that he won 90 or fewer games in the group of 202?" and the answer is around 11.6%. So I think that it is fair to say that it is not proof but it is solid evidence.Kirby wrote:8% win rate difference from a handful of human games doesn't seem like much evidence to make much of a point about a single move out of hundreds that determined the result of the game.
Of course, there are all sorts of things that confound this. Maybe Black usually makes his decision based on a situation elsewhere on the board, and that other situation is really the thing contributing to the difference in win rate. Maybe Black didn't figure out how to follow up a well until recently, so there are a lot of irrelevant games in the dataset. And so forth. But I think the difference is big enough to start with the assumption that there is something there, rather than that there is not.
(By the way, you can turn your logic around by saying that if one single move out of hundreds can change the winning probability by anything close to 8%, then it must be a pretty important decision!)
-
Gomoto
- Gosei
- Posts: 1733
- Joined: Sun Nov 06, 2016 6:56 am
- GD Posts: 0
- Location: Earth
- Has thanked: 621 times
- Been thanked: 310 times
Re: Joseki 4-4, low approach, low extension, contact variati
That's one small difference in winrate for mankind, but one giant leap for me. 
I did not claim b is the better move in all circumstances.
I propose to make b the first candidate move
In all honesty, it is a pretty significant difference (we are talking about actual winrate not an AI computed winrate here).
I took a look at the empty quadrant for starters (like shown in the diagramms), relation to board position is the next quest
I did not claim b is the better move in all circumstances.
I propose to make b the first candidate move
In all honesty, it is a pretty significant difference (we are talking about actual winrate not an AI computed winrate here).
I took a look at the empty quadrant for starters (like shown in the diagramms), relation to board position is the next quest
-
Gomoto
- Gosei
- Posts: 1733
- Joined: Sun Nov 06, 2016 6:56 am
- GD Posts: 0
- Location: Earth
- Has thanked: 621 times
- Been thanked: 310 times
Re: Joseki 4-4, low approach, low extension, contact variati
Go is a beautiful game and nobody has to choose their moves by looking at winrates if he does not like so.
It is a joy to try different things and not care about these kind of statistics. I respect and recommend to do so wholeheartedly.
But it is also perfectly viable to look at the differences in winrates and I have quite some success improving my play with this approach.
It would be stupid (I) to only look at a small difference in one postion and make final decisions on such limited evidence. But to look (II) how the winrate develops through different variations and compare this data to actual human pro games, this can give you quite some pointers, which moves are worthy to be explored in your games.
So I feel not guilty of (I) but I admit to be guilty of (II)
(a.k.a I not only enjoy playing go, I also enjoy the hunt for winrates)
It is a joy to try different things and not care about these kind of statistics. I respect and recommend to do so wholeheartedly.
But it is also perfectly viable to look at the differences in winrates and I have quite some success improving my play with this approach.
It would be stupid (I) to only look at a small difference in one postion and make final decisions on such limited evidence. But to look (II) how the winrate develops through different variations and compare this data to actual human pro games, this can give you quite some pointers, which moves are worthy to be explored in your games.
So I feel not guilty of (I) but I admit to be guilty of (II)
(a.k.a I not only enjoy playing go, I also enjoy the hunt for winrates)
-
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: Joseki 4-4, low approach, low extension, contact variati
This is a sneaky way of saying things, because you implicitly state that an 8% difference in winning probability occurs from the move, taking that as fact. This is a lot different from an AI doing playouts from A GIVEN board position. Here, we are taking a small section of the board and looking at a single move.dfan wrote:
(By the way, you can turn your logic around by saying that if one single move out of hundreds can change the winning probability by anything close to 8%, then it must be a pretty important decision!)
Taken to an extreme, you could compare a one space jump to a knight's move. If games having more knight's moves than one space jumps resulted in an 8% difference in winning across a few hundred games, does it mean that knight's move is superior to one space jump?
This is a ridiculously simplified model IMO, so I call BS.
be immersed
-
dfan
- Gosei
- Posts: 1599
- Joined: Wed Apr 21, 2010 8:49 am
- Rank: AGA 2k Fox 3d
- GD Posts: 61
- KGS: dfan
- Has thanked: 891 times
- Been thanked: 534 times
- Contact:
Re: Joseki 4-4, low approach, low extension, contact variati
If the player who made more knight's moves won 8% more games than the player who made more one space jumps over a large number of games, it would certainly cause me to think that the issue was worth looking at in greater detail.Kirby wrote: Taken to an extreme, you could compare a one space jump to a knight's move. If games having more knight's moves than one space jumps resulted in an 8% difference in winning across a few hundred games, does it mean that knight's move is superior to one space jump?
Yes, see the second paragraph in my previous comment for the very beginning of a list of issues with it. But you can learn a lot of things from simplified models, as long as you don't take them to extremes (such as the famous example of the spherical cow). For me, an 8% difference in win rate over 300 games is enough to make me think that there may be something there. Others are allowed to think differently!This is a ridiculously simplified model IMO, so I call BS.
-
Gomoto
- Gosei
- Posts: 1733
- Joined: Sun Nov 06, 2016 6:56 am
- GD Posts: 0
- Location: Earth
- Has thanked: 621 times
- Been thanked: 310 times
Re: Joseki 4-4, low approach, low extension, contact variati
(Kirby I understand your argument and do not want to contradict you with this)
Just for others reading the thread to be clear:
We use only a quarter of the board to look for a pro game sample.
The winrates in this thread are not AI computed values for a single move, but actual winrates of whole board pro game positions.
Just for others reading the thread to be clear:
We use only a quarter of the board to look for a pro game sample.
The winrates in this thread are not AI computed values for a single move, but actual winrates of whole board pro game positions.
-
Gomoto
- Gosei
- Posts: 1733
- Joined: Sun Nov 06, 2016 6:56 am
- GD Posts: 0
- Location: Earth
- Has thanked: 621 times
- Been thanked: 310 times
Re: Joseki 4-4, low approach, low extension, contact variati
And as a first hint to the whole board relations:
If we add the black hoshi stone there are 66% won games (n=15) with black b versus 46% won games with black a (n=30).
While a is played much more often by pros than b, it does show an extremly bad performance versus b in this small sample.
(But my AIs are not supporting this small sample picture )
If we add the black hoshi stone there are 66% won games (n=15) with black b versus 46% won games with black a (n=30).
While a is played much more often by pros than b, it does show an extremly bad performance versus b in this small sample.
(But my AIs are not supporting this small sample picture
)-
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: Joseki 4-4, low approach, low extension, contact variati
8% winrate difference over a few hundred games actually seems pretty big to me, comparing to other differences I’ve observed. Back when I played on OGS, I used to study fuseki by looking at kombilo/waltheri winrates over pro game databases. I wouldn’t take it as gospel that one move is better than other, but even a 1% difference over a few hundred games I would take as a valid clue a move was maybe better, and you could sometimes see how the newer pro move replaced the older one matching up with a small win rate improvement. I recall JF mentioning one pro, perhaps Hane Naoki, did some similar computer-based pattern search research on the Chinese opening (and that pre-AlphaGo this was unusual). Just as a quick example waltheri shows the slide after approach the 4-4 in Chinese opening at 45% (and 393 games, most popular) with 3-3 jump in at 54% on 179 games. Even before AlphaGo this 3-3 had become popular and considered the better move, and AIs agree.
All other things being equal, I think even a few percent difference in winrate is a useful clue a move is better. Kirby’s reservations possibly come from a belief that all other things aren’t equal: maybe there’s lots more games of Lee Sedol in the ‘b’ set than the ‘a’ set, and Lee Sedol wins a lot so you can’t say those wins are because of his choice in this position, but just because Lee Sedol’s super strong. But what if Lee Sedol wins a lot because he chooses good moves, and this is an example of a good move? Ah, but Lee Sedol wins because he is a very strong fighter and wins in middle game one counters. But maybe he wins in middle game because he builds strong positions in the opening, such as with this ‘b’ choice, that allow powerful play in later fighting. But what if it’s the losing ‘a’ set that actually has lots of Lee Sedol games? Then these arguments that were aiming to dismiss the win% difference would, if valid, actually say the “true” difference would be more were it not for the Lee Sedol bias. So without any evidence that there is a systematic bias towards one set of games, I think it’s reasonable for the a priori view to attribute the win% difference to the move difference chosen, because that is the only selected difference in the populations. Plus, isn’t the success of AlphaGo, LeelaZero etc evidence this approach works? They train their neural networks by looking at whether a single move results in a game being won or lost much later, and assumes there is a causal link. I don’t really understand stochastic gradient descent and all that jazz, but I wouldn’t be surprised if AG/LZ are learning from winrate differences of less than 1% on differential data sets of a pattern like this of less than 100 games. Learning from tiny differences, crank the handle millions of times, and you end up with a super-human player.
All other things being equal, I think even a few percent difference in winrate is a useful clue a move is better. Kirby’s reservations possibly come from a belief that all other things aren’t equal: maybe there’s lots more games of Lee Sedol in the ‘b’ set than the ‘a’ set, and Lee Sedol wins a lot so you can’t say those wins are because of his choice in this position, but just because Lee Sedol’s super strong. But what if Lee Sedol wins a lot because he chooses good moves, and this is an example of a good move? Ah, but Lee Sedol wins because he is a very strong fighter and wins in middle game one counters. But maybe he wins in middle game because he builds strong positions in the opening, such as with this ‘b’ choice, that allow powerful play in later fighting. But what if it’s the losing ‘a’ set that actually has lots of Lee Sedol games? Then these arguments that were aiming to dismiss the win% difference would, if valid, actually say the “true” difference would be more were it not for the Lee Sedol bias. So without any evidence that there is a systematic bias towards one set of games, I think it’s reasonable for the a priori view to attribute the win% difference to the move difference chosen, because that is the only selected difference in the populations. Plus, isn’t the success of AlphaGo, LeelaZero etc evidence this approach works? They train their neural networks by looking at whether a single move results in a game being won or lost much later, and assumes there is a causal link. I don’t really understand stochastic gradient descent and all that jazz, but I wouldn’t be surprised if AG/LZ are learning from winrate differences of less than 1% on differential data sets of a pattern like this of less than 100 games. Learning from tiny differences, crank the handle millions of times, and you end up with a super-human player.