Uberdude's journal

[dust]

The British Go players' synopses of their pandanet games used to be a source of much unintended hilarity for connoisseurs of Alan Partridge style reporting. Unfortunately these players have become somewhat more self-aware over the years.

Just for fun, here's my version of a template synopsis from the golden era:


My professional opponent seemed unfamiliar with the latest thinking on the mini-chinese fuseki, and by move 25 I calculated i'd drawn ahead by komi. Whilst motoring even further ahead by about 15- 20 points in the middle-game, my usual complacency set in. My opponent made a last-ditch attempt to turn round a lost game by threatening to cut my large central group, and - fatally distracted - I temporarily forgot the 17 move sequence which I'd already read protected against the cut. Despondency then set in and I somehow managed to achieve the unlikely result on dying on both sides, and losing two secure corner groups.

A shame to lose a completely won game, but full credit must go to my opponent for not giving up and taking advantage of a silly mistake to snatch victory from the jaws of defeat.

[dhu163]

if you don't know about leelabot 4d already, I highly recommend it for reviewing games. It uses monte carlo simulation, with 1 neural network (policy), and is free since april.

I made an account for it on kgs, and running on my laptop, it is already 5d.

It has serious issues with life and death (it thinks dead groups are alive, but if it tries to kill you, you are probably dead), and capture races - it's reading isn't great, and it may play out ladders that fail. Like alphago, it has a preference for settling a position. But such weaknesses just means its global judgement is very good. It values the centre more than most, and it is beautiful to watch a game where it sacrifices 3 side groups in quick succession just to build a massive centre.

I find watching it has opened my mind to more ideas, I am less dismissive of unusual moves, and appreciate that there are more good moves per position than I can see. I like to use it to analysis openings - direction, and when to fight - it is very sensitive to aji. But not so interested in its poor endgame.

I have used it to add notes to one of your games: simulation number (usually in the millions), and winning probability for a selection of moves. Up to 116

hope it's useful/interesting

[Fedya]

First off I struggled to find a slow dan game

Do you put up a custom game with your preferred time controls and find that nobody accepts the challenge?

The games I put up generally have more time than most games, and it's only 20 minutes plus byo-yomi, but people accept and then mostly don't use the time. As long as people don't complain when I actually do use mine....

[Uberdude]

Thanks for the leela analysis dhu, I had a quick look through some of the suggestions though I did keep forgetting whether coordinates were suggestions for this or the next move! The "!!!!" means current move it doesn't like though right? So for example at move 29 position your comments mean instead of my n11 turn it prefers f5 (it seems rather obsessed with this move) with 47.31 win %. Then on move 30 the comments are for black's move 31, it prefers o9 jump instead of hane, and m10 53.47 is a typo for n10 because m10 now is crazy, or was it m10 for white's last move? Will look more later.

Fedya, yes I put up an offer and also look for others'. I did get some challengers from 1 kyus but wanted to play even games.

[Uberdude]

Just making a note of something interesting I noticed in the recent commentary on Artem's win against Sonoda Yuichi. On another thread recently I said:

One other thing that I noticed I had a habit of doing (I think I have somewhat corrected it) was making overly honte net capture of a cutting stone in some big wall I would make after my opponent invaded and took some territory. Often being gote was a problem: I thought I was being "professional" and playing honte, but actually it was slow. I say "actually" with a little caution because maybe my reviewer was wrong (or it not wrong, the position was not so clear cut). Probably quite a few of those reviews were from Guo Juan 5p, who is undoubtedly very strong, but not a top pro (though does admit when she doesn't know some things), though I think some others were from Matthew Macfadyen 6d who probably has a rather overplayish and not honte-when-you-should style. Anyway the point I took away from these sort of positions is I didn't need to defend against some cutting stone plastered against my big wall from running away when such running would start a hugely advantageous fight for me. Of course as the board develops the balance of the fight might change and I might need to come back later.

One of Artem's moves that came in for criticism was exactly this sort of too slow honte net (of lower right cutting stone), he should have pressed like in the diagram (Sonoda got the correct direction and answered the net with jumping out to left of 3):

Var. 5: Instead of the geta 90, White should consider a more active play.

[Uberdude]

Also something to study: this new joseki that's cropping up in pro games a lot recently. I actually remember wondering why people didn't play this descent instead of block back when I first came across this joseki so it seems there has been a re-evaluation:

$$W
$$ . . . . , . . . |
$$ . . . . . . . . |
$$ . . . . . 4 . . |
$$ . . . . . . . . |
$$ . . . . . 1 7 . |
$$ . . . . . 6 . . |
$$ . . . . X . 3 . |
$$ . . 2 . . 5 . . |
$$ . . . . . . . . |
$$ . . . . . . . . |
$$ ----------------+

Click Here To Show Diagram Code

[go]$$W
$$ . . . . , . . . |
$$ . . . . . . . . |
$$ . . . . . 4 . . |
$$ . . . . . . . . |
$$ . . . . . 1 7 . |
$$ . . . . . 6 . . |
$$ . . . . X . 3 . |
$$ . . 2 . . 5 . . |
$$ . . . . . . . . |
$$ . . . . . . . . |
$$ ----------------+[/go]

[Shenoute]

Also something to study: this new joseki that's cropping up in pro games a lot recently. I actually remember wondering why people didn't play this descent instead of block back when I first came across this joseki so it seems there has been a re-evaluation:

$$W
$$ . . . . , . . . |
$$ . . . . . . . . |
$$ . . . . . 4 . . |
$$ . . . . . . . . |
$$ . . . . . 1 7 . |
$$ . . . . . 6 . . |
$$ . . . . X . 3 . |
$$ . . 2 . . 5 . . |
$$ . . . . . . . . |
$$ . . . . . . . . |
$$ ----------------+

Click Here To Show Diagram Code

[go]$$W
$$ . . . . , . . . |
$$ . . . . . . . . |
$$ . . . . . 4 . . |
$$ . . . . . . . . |
$$ . . . . . 1 7 . |
$$ . . . . . 6 . . |
$$ . . . . X . 3 . |
$$ . . 2 . . 5 . . |
$$ . . . . . . . . |
$$ . . . . . . . . |
$$ ----------------+[/go]

Interesting. I saw it for the first time yesterday in a game played on Tygem by Meng Tailing (video here). Black's position is not as strong as in the joseki, but white doesn't get as much points and has no ladder breaker to aim to.

[Uberdude]

Another Pandanet European Team Championship game.



I stood in for Bruno against Zoran Mutabžija (who was a 5d European Champion many decades ago but playing at 2d now) and won by a comfy 27.5 in the end although there were a few nervous moments. He started with a diagonal opening and in a pincer counter-pincer running fight I spent a long time reading out a press expecting him to push and cut but he just crawled on the 3rd line! Then he made some bizarre attach and crosscut sacrificing his pincer group and I got thick but then played 2 perhaps slow moves and he developed quickly on the rest of the board. Peeved that he didn't defend his top right corner I attacked it but it connected out on the 1st line weakening my thickness so I should have just left it unplayed as after I returned to answer his approach he nicely hassled my thickness and it ended up squirming for connection. I managed to recover somewhat and cut off 2 stones with some peepy tesujis to grow my right side which I then solidified into solid points in exchange for him making a huge double wing moyo from his lower left shimari. I was a bit nervous now and misread a way in so tenukied to cap his shimari and luckily he quickly ignored (I was in byo-yomi by now, he hadn't used much time) and then didn't try to kill me so I managed to live quite comfortably inside his moyo and from then on it was an easy endgame (with a cute tombstone tesuji he missed to capture a few stones).

[Uberdude]

Well, I've not been playing studying much, but I did go to the London Open which was fun. I only won 3 out of 7, but it was a very strong field this year and I got to play two 7ds (including Haylee) so that was a reasonable result. I ended in 11th place and was surprised to win the David Ward cup for top British (as in eligible for British Go Championship which is citizen or 5 years residence) player with so few wins. I started off with decent wins against 4d and 5d, then lost to 7d (Kim Seongjin) but gave a decent game and actually made him worry about being behind with a cool centre haengma. I played slowly though and then fell to bits and lost on time in byoyomi. Then I played Haylee, she went for territory and I didn't embarrass myself too much but she gradually pulled ahead and when her last reduction group connected to safety, I didn't harass the remaining thin group and played bad endgame she won by about 20. Next I lost to a 4d in a rather regretful game, losing by 2.5 after a slack late middlegame and several point losing time tesujis. Then lost to xhu (entered as 3d but his good results suggest Euro 4d would be correct rank) in a pretty close game with ups and downs, and then beat a (good) 2d by 1.5 after a big moyo game (which featured tenuki of 2 stones in 5-3 high approach we talked about with Bill from my game with Surin).

Here are the games with a few comments from me/opponent/Catalin Taranu.

[Fedya]

In the first game, why isn't aji keshi?

[illluck]

In the first game, why isn't aji keshi?

It was considered aji keshi until Alphago played it - seems like idea is that even though there are potentially other local moves there they aren't usually played so there's not much loss in making the exchange now. Might be losing a threat but in exchange getting that trade in now also makes black's group stronger (later white might resist instead of connect).

All this is above my level so would be happy to hear corrections/additional comments.

Edit: Uberdude - for variation at move 23 in first game, could the aji around E14 (in addition to white's emergency move at B12) be the reason?

[Uberdude]

As illluck said, AlphaGo has caused a re-evaluation of that peep: if you'd asked a year ago I would have said yes it's aji keshi and pros don't do that (actually there are a small number of examples pre-2016), but now it's a very popular pro move thanks to AlphaGo playing it vs Lee Sedol in game 2. If black later wants to tenuki from that group or do some fast-paced opening probing before adding a move, that peep can come in handy to make the group more resilient in future attacks and you might not be able to get it later, for example after white's c12 like in my game (e.g. if you peep after c12 white might just connect at b14 (though some atari problems) and it also helps cover the e13 cut in those shapes), or resist with the e14 counter peep (e.g. http://www.lifein19x19.com/forum/viewto ... 128#p86128). Another issue is if white answers not with the solid connection but tries to be faster/more efficient with g18/17 then not only does that have a downside of the push and cuts being ko threats, but black b15 becomes semi-sente (c17 wedge is a tesuji followup) which makes it harder to play the c12 invasion.

That peep is now also sometimes played in what I call the AlphaGo opening (from game 5), for example Takao Shinji vs Iyama Yuta in the Meijin final game (http://www.lifein19x19.com/forum/viewto ... 31#p212631). I did that (and he didn't resist) in my 3rd game vs Kim Seongjin 7d (photo by Marek Labos: https://www.facebook.com/photo.php?fbid ... =3&theater).

As for that f12 connection, it's the normal pro move (and usually black hasn't made the peep) so better as I thought, I must have misremembered or confused it with another situation to think the other connection was standard.

P.S. Marek's London photo album: https://www.facebook.com/profile.php?id ... =3&qsefr=1

[John Fairbairn]

I'm hoping Mike Novack might come on in this as he explains AI rather well, but if we look at the peep (B13, Game 1) first, there are about 100 non-AlphaGo games with this, going back 50 years. The side making the peep has a win rate of 51.5%. Presumably that's quite significant to a pro, but is it the sort of datum that AlphaGo is actually using?

Interestingly, during the last year there has been a trend for the peepee to make the counter-peep before answering the peep. No doubt the number of games is too small to be significant (10), but in those games the counter-peeper has a 60% win rate.

[dust]

Good games.

Off-topic, but I'd find this a bit off-putting:

https://www.facebook.com/photo.php?fbid ... =3&theater

[dfan]

When AlphaGo's policy network was first trained, it was trained to match human moves when given positions from human games, full stop. It didn't care who won the original games. Afterwards it improved its network solely through self-play. So it never used the data of how often a move worked, only how often a move was played.

From my experience using Waltheri, 51.5% is not a particularly interesting win rate for a move (I see plenty of moves that seem to be reasonably common despite having a win rate of under 45%, for example).