I used the reddit zip of 60 games as source ( Marcel Grünauer)
This is clearly Fan Tingyu 9p, in Korean p=f
this is my review, I didn't touch leelabot this time for fear it corrupts my thoughts
edit 01 Feb: have gone through using leela, and added a few variations
lets review master
-
dhu163
- Lives in gote
- Posts: 474
- Joined: Tue Jan 05, 2016 6:36 am
- Rank: UK 2d Dec15
- GD Posts: 0
- KGS: mathmo 4d
- IGS: mathmo 4d
- Has thanked: 62 times
- Been thanked: 278 times
lets review master
- Attachments
-
- 01-2016.12.29-Magist-Pan_Tingyu.sgf
- (24.42 KiB) Downloaded 2439 times
Last edited by dhu163 on Fri Feb 03, 2017 3:26 am, edited 7 times in total.
-
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: lets review master
In the variation with white o18 descent and looking at black m17 invasion aji, white has the l17 bump answer to reduce problems (though still some sente moves from centre and locally speaking it might not actually be 100% alive but is enough until black plays 20 times in a row).
-
dhu163
- Lives in gote
- Posts: 474
- Joined: Tue Jan 05, 2016 6:36 am
- Rank: UK 2d Dec15
- GD Posts: 0
- KGS: mathmo 4d
- IGS: mathmo 4d
- Has thanked: 62 times
- Been thanked: 278 times
Re: lets review master
I've edited it to show that
Game 2
Game 2
- Attachments
-
- 02-2016.12.29-Magist-Zhang_Ziliang.sgf
- (9.86 KiB) Downloaded 2437 times
Last edited by dhu163 on Fri Feb 03, 2017 3:48 am, edited 3 times in total.
-
dhu163
- Lives in gote
- Posts: 474
- Joined: Tue Jan 05, 2016 6:36 am
- Rank: UK 2d Dec15
- GD Posts: 0
- KGS: mathmo 4d
- IGS: mathmo 4d
- Has thanked: 62 times
- Been thanked: 278 times
Re: lets review master
Game 3
I find the naming of the players in this source a little suspect - the game marked as Yun ChanHee, I think I've seen named as being Lee Donghoon in two other places I think.
I find the naming of the players in this source a little suspect - the game marked as Yun ChanHee, I think I've seen named as being Lee Donghoon in two other places I think.
- Attachments
-
- 03-2016.12.29-Ding_Shixiong-Magist.sgf
- (7.77 KiB) Downloaded 2524 times
Last edited by dhu163 on Wed Feb 01, 2017 8:25 am, edited 2 times in total.
-
wolfking
- Lives in gote
- Posts: 378
- Joined: Tue Feb 18, 2014 3:33 pm
- Rank: KGS 1d
- GD Posts: 0
- KGS: wolfking6504
- Tygem: wolfking97
- Has thanked: 124 times
- Been thanked: 364 times
Re: lets review master
dhu163 wrote:Game 3
I find the naming of the players in this source a little suspect - the game marked as Yun ChanHee, I think I've seen named as being Lee Donghoon in two other places I think.
I have read somewhere that dauning(P) was a shared account by several top pros, so maybe Lee Donghoon was a first guess then someone leaked who was using the account at the time. Just my 2 cents.
Last edited by wolfking on Wed Feb 01, 2017 8:57 am, edited 1 time in total.
-
dhu163
- Lives in gote
- Posts: 474
- Joined: Tue Jan 05, 2016 6:36 am
- Rank: UK 2d Dec15
- GD Posts: 0
- KGS: mathmo 4d
- IGS: mathmo 4d
- Has thanked: 62 times
- Been thanked: 278 times
Re: lets review master
wolfking - I think you are right, as this source also shows Kim Jiseok playing 1 not 2, corroborating Haylee's statement of cyberoro news.
Master changed how we look at the Chinese fuseki with the game with Fan Tingyu, and this game drastically changes how we see the Kobayashi fuseki.
I feel that Master likes to play active moves, not moves with loose and multiple meanings - either live clearly, or sacrifice, either good shape for own area or good shape to pressure opponent area. For example the Chinese fuseki attachment, or this Kobayashi one space not two space approach, and many others. For example I don't see it jumping 2 space when pincered approaching a 3-4, which was so common a few years or more back. Also it likes to cut quite viciously sometimes, inventing many new joseki like this (vs Iyama, Gu Li final game). My guess is that this is both a matter of its style and an improvement on how to find the best move.
Its local reading seems to be very powerful. Though these are fast games, it shows huge strength in cross-cut fighting. This was said to be a weakness after the Lee Sedol match, but the game 4 errors seem unlikely now. And if it has superior global judgement, how can it be beaten? I think we must maintain some scepticism over its global judgement as that is an area where perfection is impossible and there is always room for improvement.
Also it isn't clear how strong its endgame really is, given that it happily gives up 20 points in the end and still wins. I think that it probably has several weaknesses in the endgame, especially with ko, as there is too much reading to compute and too many weird but good choices for the neural network to find the exact best lines (comparatively to its strengths vs humans). And it is not that easy to program the maths involved.
But many pros say they think top pros can give it a good no komi game. Personally I have my doubts - I think it will show more strength under more pressure. Perhaps the time settings need to be very long.
Edit: A good summary of some of its type of innovations is attach on top of 3rd line stone, and shoulder hit 3rd line stone, your influence is weaker than you think, taking sente is very important. I wouldn't say it particularly likes the outside, but that it can use it well. It has some very interesting tenukis on weakish groups, placing a lot of importance on the value of each move. But it also confirms a lot of slightly uncertain ideas pros already have about joseki/fuseki
Game 4
I'll aim to look at these in sets of 3 letting the games simmer in my mind for a few days before applying leelabot. I also repeatedly edit when I notice more things.
Also I'm pretty sure last year or so Meng Tailing played something similar to the top right vs Kang Dongyun (maybe?) and commented why does no-one play like this?
Master changed how we look at the Chinese fuseki with the game with Fan Tingyu, and this game drastically changes how we see the Kobayashi fuseki.
I feel that Master likes to play active moves, not moves with loose and multiple meanings - either live clearly, or sacrifice, either good shape for own area or good shape to pressure opponent area. For example the Chinese fuseki attachment, or this Kobayashi one space not two space approach, and many others. For example I don't see it jumping 2 space when pincered approaching a 3-4, which was so common a few years or more back. Also it likes to cut quite viciously sometimes, inventing many new joseki like this (vs Iyama, Gu Li final game). My guess is that this is both a matter of its style and an improvement on how to find the best move.
Its local reading seems to be very powerful. Though these are fast games, it shows huge strength in cross-cut fighting. This was said to be a weakness after the Lee Sedol match, but the game 4 errors seem unlikely now. And if it has superior global judgement, how can it be beaten? I think we must maintain some scepticism over its global judgement as that is an area where perfection is impossible and there is always room for improvement.
Also it isn't clear how strong its endgame really is, given that it happily gives up 20 points in the end and still wins. I think that it probably has several weaknesses in the endgame, especially with ko, as there is too much reading to compute and too many weird but good choices for the neural network to find the exact best lines (comparatively to its strengths vs humans). And it is not that easy to program the maths involved.
But many pros say they think top pros can give it a good no komi game. Personally I have my doubts - I think it will show more strength under more pressure. Perhaps the time settings need to be very long.
Edit: A good summary of some of its type of innovations is attach on top of 3rd line stone, and shoulder hit 3rd line stone, your influence is weaker than you think, taking sente is very important. I wouldn't say it particularly likes the outside, but that it can use it well. It has some very interesting tenukis on weakish groups, placing a lot of importance on the value of each move. But it also confirms a lot of slightly uncertain ideas pros already have about joseki/fuseki
Game 4
I'll aim to look at these in sets of 3 letting the games simmer in my mind for a few days before applying leelabot. I also repeatedly edit when I notice more things.
Also I'm pretty sure last year or so Meng Tailing played something similar to the top right vs Kang Dongyun (maybe?) and commented why does no-one play like this?
- Attachments
-
- 04-2016.12.29-Magist-Xie_Erhao.sgf
- (10.59 KiB) Downloaded 2454 times
-
dhu163
- Lives in gote
- Posts: 474
- Joined: Tue Jan 05, 2016 6:36 am
- Rank: UK 2d Dec15
- GD Posts: 0
- KGS: mathmo 4d
- IGS: mathmo 4d
- Has thanked: 62 times
- Been thanked: 278 times
Re: lets review master
so happy L19 is back up.
I've haven't made that much progress with my alphago review but here they are
I've haven't made that much progress with my alphago review but here they are
- Attachments
-
- 07-2016.12.29-Qiao_Zhijian-Magist.sgf
- (7.22 KiB) Downloaded 2357 times
-
- 06-2016.12.29-Li_Xiangyu-Magist.sgf
- (3.92 KiB) Downloaded 2371 times
-
- 05-2016.12.29-Yu_Zhiying-Magist.sgf
- (12.98 KiB) Downloaded 2376 times
-
dhu163
- Lives in gote
- Posts: 474
- Joined: Tue Jan 05, 2016 6:36 am
- Rank: UK 2d Dec15
- GD Posts: 0
- KGS: mathmo 4d
- IGS: mathmo 4d
- Has thanked: 62 times
- Been thanked: 278 times
Re: lets review master
mostly just a rough overview so far, but no much complicated happens this game
- Attachments
-
- 08-2016.12.29-Magist-Han_Yizou.sgf
- (8.87 KiB) Downloaded 2230 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: lets review master
I thought I would resurrect this thread, now that DeepMind have released AlphaGo's evaluations of the opening of the Master series of games as part of the AlphaGo teach opening book. Michael Redmond reviewed quite a few of these games, and would often make comments like "I wish we knew what AlphaGo would do" or wondered how the human ended up looking behind already by move 30. We can't ask AlphaGo what it thinks of arbitrary moves, but we do have its "win %" metric (viewtopic.php?f=18&t=15312 for discussion of what exactly this is) for each move played in the game (up to move 30ish) plus the move it would play. Note that once we get into deeper positions AlphaGo Teach usually only gives us the game move and AG's preferred one, and the coloured circle legend on the website of "Green = move AG would play; Purple = move human would play" seems inaccurate: for these games purple means move human or alphago played in the actual game, whilst green is suggested variation from AlphaGo. Also we don't know how close in win % other moves not shown are to the one they chose to show us, they could be very close or far worse.
For the first game I chose Master game 2 vs Zhang Zilang 2p (weak pro goratings #557), as it's fairly calm opening with Master's distinctive shoulder hits, and both Redmond and dhu reviewed it so I can compare with those. Full game at http://www.alphago-games.com/view/event ... /1/move/50. Redmond's review at https://youtu.be/LJ8d_KTKZ-Q
AG gives black a 46-47 win % on the empty board with 7.5 komi, and Zhang starts with the best of 4-4 at 47.1% (it's not clear what the error is in these figures, I think about half a % given the fluctuations of subsequent moves, also AG may choose to play a move within about 1% of the best number). White 2 is shown as black down to 46.6%, but that position can't be worse for black than what AG expects when black played move 1, so there's some fluctuation/error. dhu noted Master as white doesn't seem to fear allowing diagonal opening, but for move 3 AG says the digonal 4-4 is best at 47.0 whilst the 3-4 played is 46.6! (though unclear how significant that is). Black 5 makes small shimari, big one (high or low about 1% better). In the real game Master played 6 as small low (45.4), but teaching tool shows small high a little better (44.7), we often saw it play this when black played 5 as the big low shimari. There was a consistent lesson from AlphaGo in those games and this tool that white can afford to make a shimari for move 6 in these sort of orthodox openings rather than splitting or approaching on the top side. In this case the k17 split is 47.8, 3.1% worse than e4, quite a significant mistake! o17 approach is 46.0, just -1.3. So the modern style of approach being better than split does seem to be correct, but we just didn't realise a shimari was even better!
AG approves of black 7 modern style shimari from 4-4 (lower right 3-3 invasion just 0.1% better) and then makes its own for 8 too. At this point k17 split is 2.1 worse, splits really suck according to AG, it seems corners are gold and sides are mud not silver. Black 9 then takes the middle side point, a good double purpose move in classic opening theory, but AG says q7 shoulder hit 1% better. Master made shoulder hit on left, but AG teaching tool says the other shoulder hit was 0.5% better, this is presumably the high temperature parameter during these games to increase variety (or did Aja have a readout of multiple moves and sometimes chose the non-top one?). Zhang blocked on the 4th line (in Master game 15 Park Junghwan blocked on the 3rd line), however AG says the best move is this tenuki to split left side and the following shoulder-hit obsessed continuation is expected!
After all the ignored should hits note 21 extension allowing hane at head of 2, and that it views 22 as sente for top side (what was follow-up if ignored, q17?) which allows the double hane to make a nice encircling for white.
Anyway, black didn't realise the correct response to a shoulder hit in the opening was tenuki, so white then jumped despite the AG teach saying q13 shoulder hit immediately .4% better: this is not so significant plus when there are two important sente points AG seems to like flitting between them. Interesting though it says that (as after the initial shoulder hit) after both f14 jump or q13 shoulder hit instead black should play the d9 split (a), it really is the key point of this opening. Instead black played 13 to split the lower side (normal said Redmond, or "locally it's a good move", no comment on globally) but this was 41.0, -2.3 on d9. I think we can see a similarity here to AG's unfavourable view of k17 (either for black or white), both moves are splitting a side with strong groups of the same colour in both corners, so to build or split in such an area is playing (relatively in terms of the whole board) close to thickness, the corner groups are untroubled by such a move so it's easy to tenuki and play in more open areas of the board. Also the d9 split AG recommends has some attacking effect against the top left shoulder hit group, splitting it from the lower left. White played checking extension from the right (dhu expected from left, AG teach and Redmond no comment, my answer is black AG would not play so simply as o3 allowing kick) and black made a standard 2 space extension to make a base, giving white sente to make another shoulder hit.
There's several interesting aspects in this short exchange. The 2-space extension to make a base is a very standard idea and elicited no comment from Redmond, but AG thinks it is slack and instead wants to (yes, you've guess it!) shoulder hit the shimari, then bully it into overconcentration from the other side and tenuki. I remember in the 1st AG vs Ke Jie game Bill talked about how making 2-space extensions for a base was an ancient idea but maybe we could be wrong about its importance too often.
Another interesting detail is that AG actually prefers (by 1%) to shoulder hit the top right one move earlier (14 instead of 16), before the checking extension. At this point it expect black to block at q14 and then it would play n3 as before. I think this is not just AG MCTS random flitting, but a probe: if black answers the should hit with r13 or s12 (which is the move it recommends when it was played in the game) to place more emphasis on destroying white potential on the right side (and less thickness for top side and pressure on the shoulder hit) then it would no longer play the checking extension from the right side but from left or tenuki.
Black also blocks the 2nd shoulder hit, AG says slide to s12 0.4% better and then white would d10 extension on left. By the time white jumps for move 18 black is already down to 38%, a 10% loss! Zhang played the solid extension to p14, Redmond explained how he prefers not to wedge because it makes white stronger (if white answered extend at a black prefers not to make b-c exchange). However, I thought this was slow, and Redmond also preferred tenuki to reinforce the lower group at d. AG doesn't mention this move but recommends d10 instead, a full 3.4% better than the game move, a big mistake. It also shows an interesting continuation with black using the o3 attachment to make some extra eyes for the 2-space extension group, and then white using a tesuji to chop it off again at the expense of sacrificing a stone (as this is a AG self-play black stays around 39% win):
The game continued as below, with Master playing another shoulder hit (fairly normal human one though) to press on the lower black group. AG actually shows an invasion at n17 to be a little better, but this leads to some crazy variations I don't understand and isn't in the theme of the actual game. Black blocked, and then wedged, both very normal moves, but AG thinks both are too simple and wants to use the thinness of the lower right corner to achieve a more efficient result (in Zhang's defence he probably didn't know he was playing AlphaGo in this early game so could just be playing ok simple moves which can beat most Tygem 9ds, if he knew he was playing AG maybe he would have searched harder for these more subtle and refined moves, but still tough in 30 seconds).
Here's AG's recommend first dodge. It must think diving into the corner at p2 is sente (thought I'd like to see what happens if white ignores and played j3 seeing as black ignored that shoulder hit) and then pushes up. White then tenukis to its n17 invasion obsession and black spends a thick move at j5.
So we don't actually get so see how p2 helps black if white continues, here's my first idea (if white will obediently answer o2 at q2):
AG's variation for black's wedge though is very elegant, and 5% better! Loses something in the corner but breaks white's centre seal.
After the wedge the game continued with Master's solid connection, perhaps a bit unusual locally (m4 normal) as it gives black m3 in sente, but reduces the power of j4 cut. Black is now at 30%, and then pushes through where white ignored before, but this drops the win% to 27% on move 29, what a quick collapse!
For the first game I chose Master game 2 vs Zhang Zilang 2p (weak pro goratings #557), as it's fairly calm opening with Master's distinctive shoulder hits, and both Redmond and dhu reviewed it so I can compare with those. Full game at http://www.alphago-games.com/view/event ... /1/move/50. Redmond's review at https://youtu.be/LJ8d_KTKZ-Q
AG gives black a 46-47 win % on the empty board with 7.5 komi, and Zhang starts with the best of 4-4 at 47.1% (it's not clear what the error is in these figures, I think about half a % given the fluctuations of subsequent moves, also AG may choose to play a move within about 1% of the best number). White 2 is shown as black down to 46.6%, but that position can't be worse for black than what AG expects when black played move 1, so there's some fluctuation/error. dhu noted Master as white doesn't seem to fear allowing diagonal opening, but for move 3 AG says the digonal 4-4 is best at 47.0 whilst the 3-4 played is 46.6! (though unclear how significant that is). Black 5 makes small shimari, big one (high or low about 1% better). In the real game Master played 6 as small low (45.4), but teaching tool shows small high a little better (44.7), we often saw it play this when black played 5 as the big low shimari. There was a consistent lesson from AlphaGo in those games and this tool that white can afford to make a shimari for move 6 in these sort of orthodox openings rather than splitting or approaching on the top side. In this case the k17 split is 47.8, 3.1% worse than e4, quite a significant mistake! o17 approach is 46.0, just -1.3. So the modern style of approach being better than split does seem to be correct, but we just didn't realise a shimari was even better!
AG approves of black 7 modern style shimari from 4-4 (lower right 3-3 invasion just 0.1% better) and then makes its own for 8 too. At this point k17 split is 2.1 worse, splits really suck according to AG, it seems corners are gold and sides are mud not silver. Black 9 then takes the middle side point, a good double purpose move in classic opening theory, but AG says q7 shoulder hit 1% better. Master made shoulder hit on left, but AG teaching tool says the other shoulder hit was 0.5% better, this is presumably the high temperature parameter during these games to increase variety (or did Aja have a readout of multiple moves and sometimes chose the non-top one?). Zhang blocked on the 4th line (in Master game 15 Park Junghwan blocked on the 3rd line), however AG says the best move is this tenuki to split left side and the following shoulder-hit obsessed continuation is expected!
After all the ignored should hits note 21 extension allowing hane at head of 2, and that it views 22 as sente for top side (what was follow-up if ignored, q17?) which allows the double hane to make a nice encircling for white.
Anyway, black didn't realise the correct response to a shoulder hit in the opening was tenuki, so white then jumped despite the AG teach saying q13 shoulder hit immediately .4% better: this is not so significant plus when there are two important sente points AG seems to like flitting between them. Interesting though it says that (as after the initial shoulder hit) after both f14 jump or q13 shoulder hit instead black should play the d9 split (a), it really is the key point of this opening. Instead black played 13 to split the lower side (normal said Redmond, or "locally it's a good move", no comment on globally) but this was 41.0, -2.3 on d9. I think we can see a similarity here to AG's unfavourable view of k17 (either for black or white), both moves are splitting a side with strong groups of the same colour in both corners, so to build or split in such an area is playing (relatively in terms of the whole board) close to thickness, the corner groups are untroubled by such a move so it's easy to tenuki and play in more open areas of the board. Also the d9 split AG recommends has some attacking effect against the top left shoulder hit group, splitting it from the lower left. White played checking extension from the right (dhu expected from left, AG teach and Redmond no comment, my answer is black AG would not play so simply as o3 allowing kick) and black made a standard 2 space extension to make a base, giving white sente to make another shoulder hit.
There's several interesting aspects in this short exchange. The 2-space extension to make a base is a very standard idea and elicited no comment from Redmond, but AG thinks it is slack and instead wants to (yes, you've guess it!) shoulder hit the shimari, then bully it into overconcentration from the other side and tenuki. I remember in the 1st AG vs Ke Jie game Bill talked about how making 2-space extensions for a base was an ancient idea but maybe we could be wrong about its importance too often.
Another interesting detail is that AG actually prefers (by 1%) to shoulder hit the top right one move earlier (14 instead of 16), before the checking extension. At this point it expect black to block at q14 and then it would play n3 as before. I think this is not just AG MCTS random flitting, but a probe: if black answers the should hit with r13 or s12 (which is the move it recommends when it was played in the game) to place more emphasis on destroying white potential on the right side (and less thickness for top side and pressure on the shoulder hit) then it would no longer play the checking extension from the right side but from left or tenuki.
Black also blocks the 2nd shoulder hit, AG says slide to s12 0.4% better and then white would d10 extension on left. By the time white jumps for move 18 black is already down to 38%, a 10% loss! Zhang played the solid extension to p14, Redmond explained how he prefers not to wedge because it makes white stronger (if white answered extend at a black prefers not to make b-c exchange). However, I thought this was slow, and Redmond also preferred tenuki to reinforce the lower group at d. AG doesn't mention this move but recommends d10 instead, a full 3.4% better than the game move, a big mistake. It also shows an interesting continuation with black using the o3 attachment to make some extra eyes for the 2-space extension group, and then white using a tesuji to chop it off again at the expense of sacrificing a stone (as this is a AG self-play black stays around 39% win):
The game continued as below, with Master playing another shoulder hit (fairly normal human one though) to press on the lower black group. AG actually shows an invasion at n17 to be a little better, but this leads to some crazy variations I don't understand and isn't in the theme of the actual game. Black blocked, and then wedged, both very normal moves, but AG thinks both are too simple and wants to use the thinness of the lower right corner to achieve a more efficient result (in Zhang's defence he probably didn't know he was playing AlphaGo in this early game so could just be playing ok simple moves which can beat most Tygem 9ds, if he knew he was playing AG maybe he would have searched harder for these more subtle and refined moves, but still tough in 30 seconds).
Here's AG's recommend first dodge. It must think diving into the corner at p2 is sente (thought I'd like to see what happens if white ignores and played j3 seeing as black ignored that shoulder hit) and then pushes up. White then tenukis to its n17 invasion obsession and black spends a thick move at j5.
So we don't actually get so see how p2 helps black if white continues, here's my first idea (if white will obediently answer o2 at q2):
AG's variation for black's wedge though is very elegant, and 5% better! Loses something in the corner but breaks white's centre seal.
After the wedge the game continued with Master's solid connection, perhaps a bit unusual locally (m4 normal) as it gives black m3 in sente, but reduces the power of j4 cut. Black is now at 30%, and then pushes through where white ignored before, but this drops the win% to 27% on move 29, what a quick collapse!
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: lets review master
@Uberdude
Many thanks for that detailed review.
In his 21st Century Go set Go Seigen anticipates AlphaGo's aversion to early wedges on the side and the early 4th line shoulder blows.
Go Seigen was not always consistent about the side wedges, nor did he seem to favor the shoulder blows as often as Master. Zero seems more territory oriented than Go Seigen, but it is easy to imagine Master as a reincarnation of Go Seigen. 
Many thanks for that detailed review.
In his 21st Century Go set Go Seigen anticipates AlphaGo's aversion to early wedges on the side and the early 4th line shoulder blows.
Go Seigen was not always consistent about the side wedges, nor did he seem to favor the shoulder blows as often as Master. Zero seems more territory oriented than Go Seigen, but it is easy to imagine Master as a reincarnation of Go Seigen. 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.
-
lightvector
- Lives in sente
- Posts: 759
- Joined: Sat Jun 19, 2010 10:11 pm
- Rank: maybe 2d
- GD Posts: 0
- Has thanked: 114 times
- Been thanked: 916 times
Re: lets review master
Uberdude - helpful review!
Unlike some of the other variations in the file with highly skewed winrates, this is one of the positions where I don't have even a slight sense of black as being behind. I'm curious, if you didn't know AlphaGo's evaluation at this point, would you have any preference for white in the final diagram you showed, and if so, why?
Unlike some of the other variations in the file with highly skewed winrates, this is one of the positions where I don't have even a slight sense of black as being behind. I'm curious, if you didn't know AlphaGo's evaluation at this point, would you have any preference for white in the final diagram you showed, and if so, why?
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: lets review master
lightvector wrote:Uberdude - helpful review!
Unlike some of the other variations in the file with highly skewed winrates, this is one of the positions where I don't have even a slight sense of black as being behind. I'm curious, if you didn't know AlphaGo's evaluation at this point, would you have any preference for white in the final diagram you showed, and if so, why?
Well, I suppose that a 30% estimate corresponds roughly to being around 20-25 pts. behind (after
) by conventional human estimates. Out of curiosity, I evaluated that position with my simple, biased evaluation function, for an estimate of 26+ pts. for White. I usually figure the central bias at around 4 pts. at that stage of the game, which yields an estimate of around 22 pts. Surprise, surprise! 
Why the difference? Short answer: Black is too concentrated and too low -- at least, by how my function figures things.
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.
-
WindCaliber
- Dies with sente
- Posts: 75
- Joined: Sun Jul 02, 2017 7:21 pm
- Rank: SDK
- GD Posts: 0
- Has thanked: 10 times
- Been thanked: 8 times
Re: lets review master
Uberdude wrote:Instead black played 13 to split the lower side (normal said Redmond, or "locally it's a good move", no comment on globally) but this was 41.0, -2.3 on d9. I think we can see a similarity here to AG's unfavourable view of k17 (either for black or white), both moves are splitting a side with strong groups of the same colour in both corners, so to build or split in such an area is playing (relatively in terms of the whole board) close to thickness, the corner groups are untroubled by such a move so it's easy to tenuki and play in more open areas of the board. Also the d9 split AG recommends has some attacking effect against the top left shoulder hit group, splitting it from the lower left.
...
There's several interesting aspects in this short exchange. The 2-space extension to make a base is a very standard idea and elicited no comment from Redmond, but AG thinks it is slack and instead wants to (yes, you've guess it!) shoulder hit the shimari, then bully it into overconcentration from the other side and tenuki. I remember in the 1st AG vs Ke Jie game Bill talked about how making 2-space extensions for a base was an ancient idea but maybe we could be wrong about its importance too often.
I had an exchange on YT comments a couple of months ago that I thought was interesting and relevant to this:
·sev· wrote:Inseong 8d had a lecture on Alphago moves in which he pointed out that the evolution from the 70's 80's Japanese style to the 90's korean style is a matter of tightening moves and making them less slack and more forceful. He continued this narrative by suggesting that the attach instead of slide is of a similar logic, and that alphago is showing many attachments to move one line closer as a sort of improvement upon the improvement.
to which I responded that the (mostly early) wariuchi might also be an example of this. This does seem to be a sente thing. A splitting move between to shimari does not necessarily require a response, as that would just be playing close to your thickness, as you mentioned.
-
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: lets review master
lightvector wrote:Unlike some of the other variations in the file with highly skewed winrates, this is one of the positions where I don't have even a slight sense of black as being behind. I'm curious, if you didn't know AlphaGo's evaluation at this point, would you have any preference for white in the final diagram you showed, and if so, why?
Yes (though it's hard to ensure my current knowledge of AG's winrate doesn't even subconsciously affect my opinion). I tend to judge positions like this more as a sum of mistake deltas than counting them. I definitely thought black's solid extension and then push through (19 and 29) at the top right were too slow and solid and a mistake, see tewari below. As for blocking both shoulder hits on the 4th line rather than crawl that's certainly a solid/thick way of playing but I can't really say bad with confidence. That tenuki is often the best response to a shoulder hit in the opening is not something I was aware of back when these games were played, but now that it's the standard AG Zero vs AG Master opening (http://www.alphago-games.com/view/event ... /0/move/16) I can nudge that 'solid' towards 'too solid and potentially bad'. h3 slackly making a base passed me by, though I can see similarities with the modern style of attaching to the shimari in orthodox wedge openings rather than simply making the extension. k17 being slow and focused on one side of the board is certainly something I could feel, but without AG's judgement I'd class it more as a style issue "pros and cons, possible to play this way if you are confident dealing with moyos later".
1 is much too slow in the opening. 2 for 3 is slight minus (e.g. losing things like a/b attachments) and slightly helps black towards corner but there is still plenty to do there (like c AG played in game), much less bad than 1. 4 is kinda suspicious because black should tenuki it and you don't know without looking at the whole board if it has much value for white, but if black answers at 5 that's a nice exchange for white. Jumping to 6 is locally sensible, whether it's worth spending a move on depends on the whole board but treating the 4-5-6 exchange as a gote move for white you make if you want some centre influence seems fine for white to me. And white has a follow up to somewhat close the right side so black might even spend a move around d like he did in the game, which does make some side points.
With black finishing at d here's another position with the same localy tally of black adding one move to his shimari. It's a sensible move locally we see fairly often, but would black wanted to have played it so early in this opening? It's like white has then played the sequence a-f which could be aji keshi (maybe you might want to invade instead of reduce), but black's answers are rather soft and leave the ko aji at g dhu mentioned. If black wanted to avoid that and connect on the side he could have played f as clamp below a for example, or simply crawled on the 3rd line instead of b.