I tried analysing AlphaGo self-play game #38. It's a rather meandering and inconclusive for a review: some bits I think I understand what it's doing, others are still mysterious.
Uberdude's journal
-
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: Uberdude's journal
- Attachments
-
- AG vs AG - G38 - Uberdude commentary.sgf
- (32.54 KiB) Downloaded 2127 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: Uberdude's journal
I learnt from a recent pro game that the meaning of the B3 probe is to make B8 sente (as B4 is threatened)
-
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: Uberdude's journal
nice stuff,
it seems better for me to create a new sgf for comments, rather that add to your large file.
isn't the Q13 problem a seriously bugging question. I don't think I can believe that a ladder breaker is sente. And running leela for an hour didn't change its opinion that if W Q13, then 60% to W, if not then 40%
it seems better for me to create a new sgf for comments, rather that add to your large file.
isn't the Q13 problem a seriously bugging question. I don't think I can believe that a ladder breaker is sente. And running leela for an hour didn't change its opinion that if W Q13, then 60% to W, if not then 40%
- Attachments
-
- agag38 uberdude, dhu edits.sgf
- (8.22 KiB) Downloaded 2040 times
-
John Fairbairn
- Oza
- Posts: 3724
- Joined: Wed Apr 21, 2010 3:09 am
- Has thanked: 20 times
- Been thanked: 4672 times
Re: Uberdude's journal
Andrew: It seems to me a lot of the surprise melts away if we make the surely not unreasonable supposition that AlphaGo makes mistakes, too. Top human pros seem united in thinking they are something like four stones away from perfect play, which would mean AG must be a good three stones away - plenty of scope for error.
Furthermore, in self-play it is making errors on both sides (i.e. not always exploiting mistakes). And conceivably the program, if the same version, is making the same errors from both sides of the board.
The fact that a human can't exploit an AG error doesn't mean it's not an error.
The latest issue of Go World gives huge space to the Future of Go Summit, with discussion led by Ohashi Hirofumi, O Meien and Ko Reibun (plus game commentaries by Ohashi). The cover photo of a despairing Ke Jie will surely go down in history.
One interesting comment by O Meien factors into your own comments. Ohashi had just made a comment about the frequency of the 3-3 invasion increasing in self play games (and Ke Jie playing it more often in his recent run). O said he initially thought it was a 場合の手 (here's your cue, Bill!), or move dictated by special circumstances. But looking at the self play games, the much greater frequency suggested the opposite of it being a special case move, and indeed it was perhaps more likely that NOT playing the invasion was the 場合の手. And Ko Reibun, who apparently spoke to the Google team, indicated that the change was something engineered between different versions of the program.
My speculation: is an early 3-3 enough to tilt the game (at pro level) in the invader's favour? After all, it's not just about splitting territory and influence, the usual way of assessing it. It also settles a significant portion of the game (both in area and time) and acts as a probe - two factors a pro is capable of exploiting, and a Monte Carlo program even more so?
Furthermore, in self-play it is making errors on both sides (i.e. not always exploiting mistakes). And conceivably the program, if the same version, is making the same errors from both sides of the board.
The fact that a human can't exploit an AG error doesn't mean it's not an error.
The latest issue of Go World gives huge space to the Future of Go Summit, with discussion led by Ohashi Hirofumi, O Meien and Ko Reibun (plus game commentaries by Ohashi). The cover photo of a despairing Ke Jie will surely go down in history.
One interesting comment by O Meien factors into your own comments. Ohashi had just made a comment about the frequency of the 3-3 invasion increasing in self play games (and Ke Jie playing it more often in his recent run). O said he initially thought it was a 場合の手 (here's your cue, Bill!), or move dictated by special circumstances. But looking at the self play games, the much greater frequency suggested the opposite of it being a special case move, and indeed it was perhaps more likely that NOT playing the invasion was the 場合の手. And Ko Reibun, who apparently spoke to the Google team, indicated that the change was something engineered between different versions of the program.
My speculation: is an early 3-3 enough to tilt the game (at pro level) in the invader's favour? After all, it's not just about splitting territory and influence, the usual way of assessing it. It also settles a significant portion of the game (both in area and time) and acts as a probe - two factors a pro is capable of exploiting, and a Monte Carlo program even more so?
-
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: Uberdude's journal
Isn't it more that it guarantees that b1 will be sente, because if you make the hanging connection on the outside first then 1) black may atari in the corner, and more importantly 2) if black answers on the outside then when you play b1 black may tenuki, or push at b8 first. If you tenuki in this case then white's follow-up is capturing the 2-2 stone, certainly big but the corner isn't all dead, versus if you b3 first then then black can't tenuki or the 3 cutting stones get captured. There's also the argument that with the hanging connection first you are more willing to tenuki the b1 hane because the hanging connection then becomes an unnecessary move if you sacrifice the 2-2 stone. This is of course only valid if black is not going to play the tombstone exchange and is afraid of the ko.dhu163 wrote:I learnt from a recent pro game that the meaning of the B3 probe is to make B8 sente (as B4 is threatened)
5 is still sente here (locally speaking for corner shape, ignoring white lower group not safe) If black ignores 3 he could push at 4: if white answers then 6 lives more comfortably White's follow-up in corner is only capture 2-2 stone, black's ok to come out at 8, stone at 1 now unnecessary Or if black tenukis, then next a to live (leaving throw in ko for white) or crawl at b but then white gets the nice atari and it's a lot of 2nd line crawling to safety: White 5 tries to kill all, but 6 is still ko Whereas if we omit 1-2 exchange and play direct cut, of course black can't crawl or tenuki
-
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: Uberdude's journal
yes, to clarify, I meant comparing between A3 and B4 responses, B4 leaves a ko, but A3 leaves C8. As I couldn't initially understand why sometimes alphago plays B4 - I couldn't see the negative side. As I would be surprised to see B tenuki the atari after B4 in this situation where B is cutting aggressively, so I didn't think that was the difference here.
I think W will not descend for tombstone as B will fight with B4, and W's shape isn't great, but rather W will squeeze immediately and W is happy.
Also we wonder why W doesn't atari B1 immediately, presumably it is to leave it as a ko threat in some variations, though it's not even that clear to me in the game that the atari is urgent. It becomes important as it is no longer sente when the 3 stones are captured, and helps W in the ko at the bottom right later.
I think W will not descend for tombstone as B will fight with B4, and W's shape isn't great, but rather W will squeeze immediately and W is happy.
Also we wonder why W doesn't atari B1 immediately, presumably it is to leave it as a ko threat in some variations, though it's not even that clear to me in the game that the atari is urgent. It becomes important as it is no longer sente when the 3 stones are captured, and helps W in the ko at the bottom right later.
-
John Fairbairn
- Oza
- Posts: 3724
- Joined: Wed Apr 21, 2010 3:09 am
- Has thanked: 20 times
- Been thanked: 4672 times
Re: Uberdude's journal
Here are a few more notes from the Go World articles, in my own words and in no special order.
1. It seems the idea of an uchikomi series has been put into cold storage and the pros believe AG has effectively retired from playing humans, but it will continue doing go research. It seems AG will not release details of AG's analysis - too big a project.
2. O Meien especially doesn't seem to think we can learn to play like AG. Its play is just too different. He notes that go is above all a game of probes, but this is difficult to handle for humans whereas AG has no problems with this and can tenuki more (a point Andrew noted above). Similarly, AG is happy to make huge trades which a human may see but wouldn't dare try. AG is superior at evaluating trades and has no fear factor, he notes. He adds that he understand the results of what AG does but not the process.
3. Kato Hideki (DeepZen developer) says AG is not all that great at reading (it looks at 10,000 moves a second but compare that to 200 million in chess). It copes with this by simplifying into thickness. This is similar to the point Andrew made about AG liking to make two eyes quickly, but O has a slightly different slant - he considers AG is putting emphasis on being able to connect (which is what is meant by thickness here - solidity), which is a different way of ensuring safety. AG is apparently using linkability to prune the lookahead tree.
4. Pros seem to be coming round to the view that 7.5 komi is a mite too much.
5. Of the three big AI machines, DeepZen is the one that humans will find easiest to imitate (and it is still going great guns in the Japanese National Squad training games, which will apparently last till end December).
1. It seems the idea of an uchikomi series has been put into cold storage and the pros believe AG has effectively retired from playing humans, but it will continue doing go research. It seems AG will not release details of AG's analysis - too big a project.
2. O Meien especially doesn't seem to think we can learn to play like AG. Its play is just too different. He notes that go is above all a game of probes, but this is difficult to handle for humans whereas AG has no problems with this and can tenuki more (a point Andrew noted above). Similarly, AG is happy to make huge trades which a human may see but wouldn't dare try. AG is superior at evaluating trades and has no fear factor, he notes. He adds that he understand the results of what AG does but not the process.
3. Kato Hideki (DeepZen developer) says AG is not all that great at reading (it looks at 10,000 moves a second but compare that to 200 million in chess). It copes with this by simplifying into thickness. This is similar to the point Andrew made about AG liking to make two eyes quickly, but O has a slightly different slant - he considers AG is putting emphasis on being able to connect (which is what is meant by thickness here - solidity), which is a different way of ensuring safety. AG is apparently using linkability to prune the lookahead tree.
4. Pros seem to be coming round to the view that 7.5 komi is a mite too much.
5. Of the three big AI machines, DeepZen is the one that humans will find easiest to imitate (and it is still going great guns in the Japanese National Squad training games, which will apparently last till end December).
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: Uberdude's journal
I have responded to Uberdude in This 'n' That: https://www.lifein19x19.com/viewtopic.p ... 91#p221591 . Now let me reply to John. 
) AlphaGo still uses Monte Carlo rollouts as part of its evaluation, and for that reason may put some premium on being able to connect.
Oh, wait! I live near Berkeley, California. Maybe I can find a psychic to channel Go Seigen! 
I think, in the relatively near future, we will see young players, especially, imitate AlphaGo and other programs, as other programs get better than humans. Even at my level I am sensitive to the drops in local temperature, but I would have trouble handling the remaining stones that lack a base. Young pros would not have that problem.John Fairbairn wrote:O Meien especially doesn't seem to think we can learn to play like AG. Its play is just too different. He notes that go is above all a game of probes, but this is difficult to handle for humans whereas AG has no problems with this and can tenuki more (a point Andrew noted above).
Go Seigen, anyone? Again, like Go Seigen did in his day, young pros could, and I believe, would develop that judgement and the skills to go with it. The next decade or two will, I think, be very exciting for the development of go.Similarly, AG is happy to make huge trades which a human may see but wouldn't dare try. AG is superior at evaluating trades and has no fear factor, he notes. He adds that he understand the results of what AG does but not the process.
Not just to prune the tree, I think. Long ago I surmised that Monte Carlo based programs liked to play in the center because connecting to other stones was the easiest way to live by random play. (Large dragons never die.Kato Hideki (DeepZen developer) says AG is not all that great at reading (it looks at 10,000 moves a second but compare that to 200 million in chess). It copes with this by simplifying into thickness. This is similar to the point Andrew made about AG liking to make two eyes quickly, but O has a slightly different slant - he considers AG is putting emphasis on being able to connect (which is what is meant by thickness here - solidity), which is a different way of ensuring safety. AG is apparently using linkability to prune the lookahead tree.
) AlphaGo still uses Monte Carlo rollouts as part of its evaluation, and for that reason may put some premium on being able to connect.Maybe so, but I think that AlphaGo is showing us the future of go, and the relatively near future, at that. It's too bad that we don't have Go Seigen commenting on AlphaGo. He might say, I told you so.Of the three big AI machines, DeepZen is the one that humans will find easiest to imitate (and it is still going great guns in the Japanese National Squad training games, which will apparently last till end December).
Oh, wait! I live near Berkeley, California. Maybe I can find a psychic to channel 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.
-
Calvin Clark
- Lives in gote
- Posts: 426
- Joined: Thu Aug 13, 2015 8:43 am
- GD Posts: 0
- Has thanked: 186 times
- Been thanked: 191 times
Re: Uberdude's journal
A game of probes? An interesting viewpoint. Is this a common one expressed by other pros?John Fairbairn wrote: 2. O Meien especially doesn't seem to think we can learn to play like AG. Its play is just too different. He notes that go is above all a game of probes, but this is difficult to handle for humans whereas AG has no problems with this and can tenuki more (a point Andrew noted above).
It is notable that a pro would feel that AlphaGo is too far ahead, but I gather quite a few do. It echos the sentiment that I think many of us amateurs share that it's hard to learn from pro games (especially some of the more modern violent games) until one's reading is already approaching pro level. I am reminded of advice often given along the lines of: study the more accessible games from this pro (e.g., Otake, Takagawa, Chang Hao) not that pro whose games are hard for anyone to understand (e.g., Go Seigen or many current young pros).
Many Kahneman System 1 surfaces a lot of moves my brain somehow notices from replaying many pro games, but many of them I wouldn't be able to handle tactically or even strategically. Is playing them, then, cargo cult go? Maybe.
This is the conversation that goes on in my head:
Logical Brain: Look, Intuitive Brain. You have one job and only one job: to suggest moves that I wouldn't otherwise consider. You don't actually make the moves because you make basic mistakes that get us in trouble, so don't let me catch you doing that.
Intuitive Brain: Shut up. I'm stronger than you.
Logical Brain: Only today. Not forever.
It will be interesting see how pro's own imitation of AlphaGo differs from the way previous generations of pros learned from their masters.
-
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: Uberdude's journal
Time for a little update: I won the Isle of Man Go festival main tournament (beautiful place btw, highly recommended). The competition wasn't so tough (only 2 dans and below under the stricter new British rank rules) but some games were uncomfortably close (e.g. I forgot to gain a liberty in sente in an ~90 point semeai so lost by 1 liberty in my first game but managed to catch up, and 2nd game was also losing but won by 1.5 in the end) so I need to play better in the British title match in 2 weeks.
Anyway, in my 3rd game I tried out AlphaGo's double attachment from the 1st self play game (though Alex had pincered my approach on the left side instead of back off to a so maybe the premise of counter hane to develop there was less valid). 13 was maybe too soft, but Alex said if hane that side he didn't like my 2 counter hanes inside to settle there but I wasn't actually so keen on that thinking it was cramped. Interestingly Michael Redmond said in his recent review of the AlphaGo game that he wanted to hane there too (AG played inside hane) and wasn't sure how white would show that was not so good for black.
- 20170803_173247_crop_545x561.jpg (106.32 KiB) Viewed 27133 times
-
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:
Re: Uberdude's journal
I would think that's a too deterministic view on how AI works but I'm not a specialist myself. (Agree with the rest of what you said)John Fairbairn wrote: And conceivably the program, if the same version, is making the same errors from both sides of the board.
-
dfan
- Gosei
- Posts: 1598
- 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: Uberdude's journal
The hypothesis seems reasonable to me. Say that AlphaGo undervalues influence. As White it might play suboptimal moves that allow Black to gain an unreasonable amount of influence, thinking that it's not important; but then as Black it would not punish those moves by playing the influence-gaining response, for the same reason.Knotwilg wrote:I would think that's a too deterministic view on how AI works but I'm not a specialist myself. (Agree with the rest of what you said)John Fairbairn wrote: And conceivably the program, if the same version, is making the same errors from both sides of the board.
Ideally reinforcement learning would take care of this problem eventually, but it can be hard when you're just playing yourself and there's no one who plays in a different style to show you the error of your ways.
-
Fedya
- Lives in gote
- Posts: 603
- Joined: Tue Apr 20, 2010 8:21 pm
- Rank: 6-7k KGS
- GD Posts: 0
- Has thanked: 43 times
- Been thanked: 139 times
Re: Uberdude's journal
I hope you didn't fall off the island.
(Seriously, looks like a beautiful place to visit.)
I also agree with dfan in his comment about how are you supposed to improve if there's nobody to show you the error of your ways.
(Seriously, looks like a beautiful place to visit.)
I also agree with dfan in his comment about how are you supposed to improve if there's nobody to show you the error of your ways.
-
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
-
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: Uberdude's journal
Something amusing I noticed today: my play against Pavol Lisy earlier this year can't have been too bad as Sonoda Yuichi 9p just played pretty much the same in the Kansai Kiin championship yesterday! http://www.go4go.net/go/games/sgfview/69946
And Sonoda Yuichi 9p (white) vs Kiyonari Tetsuya 9p yesterday, the key difference is black kept on pushing instead of slide to the 2nd line, and Sonoda didn't surround the group so early (it's not sente) so black didn't have a change to come up like Pavol did with 27. Also when Catalin Taranu reviewed my game he criticised the direction of my checking extension as being from the thickness, Sonoda does the same direction though it's true his moyo to the left is more developed with the marked stones so there's more reason to do it in his game.
That Black answered 14 immediately suggests he may have thought peeping in the corner immediately white may do well do resist with the attachment at a in which case white may sacrifice the corner and gobble up the lower side. I did consider this in my game but the cutting point of 12 was not defended yet which made it harder, I think Kiyonari's earlier exchange of 11 for 12 actually makes those stones heavier so makes this exchange easier to consider for white. If white is going to extend at 16 immediately though it doesn't really matter, but white did have a choice to play something on the right at that point to prevent the peep. Sonoda didn't spend a move at a after 23 like I did, so can stay ahead of black with the pushing on the right side. That he also went for the shoulder hit of 24 rather than something softer validates my view it was not too thin but a reasonable pressing move. In the end Sonoda does enclose at 38 which also helps deal with the bad aji of the 31 stone. The lower group ends up being a seki. But my lesson is I did indeed play this move too early, my move 26 above should have been jump or press on the top side as closing wasn't sente so too slow.
And Sonoda Yuichi 9p (white) vs Kiyonari Tetsuya 9p yesterday, the key difference is black kept on pushing instead of slide to the 2nd line, and Sonoda didn't surround the group so early (it's not sente) so black didn't have a change to come up like Pavol did with 27. Also when Catalin Taranu reviewed my game he criticised the direction of my checking extension as being from the thickness, Sonoda does the same direction though it's true his moyo to the left is more developed with the marked stones so there's more reason to do it in his game.
That Black answered 14 immediately suggests he may have thought peeping in the corner immediately white may do well do resist with the attachment at a in which case white may sacrifice the corner and gobble up the lower side. I did consider this in my game but the cutting point of 12 was not defended yet which made it harder, I think Kiyonari's earlier exchange of 11 for 12 actually makes those stones heavier so makes this exchange easier to consider for white. If white is going to extend at 16 immediately though it doesn't really matter, but white did have a choice to play something on the right at that point to prevent the peep. Sonoda didn't spend a move at a after 23 like I did, so can stay ahead of black with the pushing on the right side. That he also went for the shoulder hit of 24 rather than something softer validates my view it was not too thin but a reasonable pressing move. In the end Sonoda does enclose at 38 which also helps deal with the bad aji of the 31 stone. The lower group ends up being a seki. But my lesson is I did indeed play this move too early, my move 26 above should have been jump or press on the top side as closing wasn't sente so too slow.