Life In 19x19

LOL! Andrew, you bought a nice new machine with a good graphics card so that you could restrict you opponent to minimal playouts and exploit known weaknesses?

Hey, I just used the defaults in Lizzie! And actually I think 1000 playouts is not so small: some guy on reddit is doing tests reducing LZ's playouts to get about even results vs Leela 0.11 (6d+) with unlimited thinking time and is currently down to only 55 playouts. Plus when I was reviewing my tournament games with my laptop its 1 playout instinct was often better than our moves, and with a few hundred seemed plenty stronger than us. But yes I am smarter than Leela 0.11 in knowing LZ has ladder problems so intentionally gave it opportunities to mess up.

My game against Shi Yue 9p from the 5th World Collegiate Weichi Championships / Changqi Cup in Cambridge last week (record a bit wrong at end but yes I did blunder die; had to hurry up as last to finish).



Pics: https://www.facebook.com/photo.php?fbid ... ater&ifg=1

Uberdude wrote:My game against Shi Yue 9p from the 5th World Collegiate Weichi Championships / Changqi Cup in Cambridge last week (record a bit wrong at end but yes I did blunder die; had to hurry up as last to finish)

Nice game! You seem to be a few points ahead if you live on the left.

How did the other strong amateurs do in the handicap simultaneous games?

sorin wrote: Nice game! You seem to be a few points ahead if you live on the left.

Yes, it was pretty painful losing all those points in the lower right (too much fighting spirit with g4, and not defending at s3 when I had the chance) but I had hung on to just enough. At the end I overlooked that making the e9 d9 exchange meant c6 was no longer sente so I had to live differently. I don't really know why I kicked and invaded at h17 seeing as I expected his h16 tesuji and had no good answer planned; took me a while to get into the game. The centre fighting was fun, seeing as Shi Yue is one of the strongest fighters in the world I wanted to fight against him rather than play some passive "preserve my handicap" style. I reviewed a bit with LZ, it said I basically maintained a >90% winrate all the way to the blunder, but winrate is not so useful for reviewing handicap games as in the opening 4 handicaps is ~50 points lead and given as 96% win, whereas nearer the end of the game 96% winrate could mean a sure lead of 3 points. Really what I want is a points lead graph showing my gradual (or not for big mistakes!) erosion of the 4 stones advantage as the game progresses. There is a "handicap-instead-of-winrate" setting in the lizzie config, I'll give this a go later.

sorin wrote: How did the other strong amateurs do in the handicap simultaneous games?

I only know Daniel Hu lost on 3 stones to Tuo Jiaxi, and Amir Fragman lost to Mi Yuting (don't know handi). Here you can see Laurent Heiser's game in progress

I won the first game of the 2018 British Go Championship yesterday against Sam Aitken 4d. A little bit of AI influence (press bigger than double approach to 4-4) but no modern 3-3 invasions and then a fairly traditional looking game. I did better in opening and early middle-game, messed up in middlegame and was behind losing a bunch of stones. Some sloppy endgame play from both of us meant we each had chances but in the end I came out ahead by (my now traditional title-match score of) 2.5 points. (o11 connect is losing move if I connect at p11 (he should peep inside to get the 1st line sente yose), then o13 is losing move as p13 wedge works due to me not playing p11 and is a few points better.)



Commentary by Matthew Macfadyen 6d:

Some comments.



LZ's winrate of the game with move numbers of key moments marked:

An LZ review of an old OGS game of mine which copied a pro game and I diverged when I successfully identified the pros mistake (and my opponent then kindly made lots more).

I was the British player at the 13th Korea Prime Minister's Cup in Buan, South Korea 2 weeks ago. I had a good time meeting old and new friends and played fairly decent Go given my worries about the short time limits (40 mins and 30 seconds 3 times) and ended with a decent 18th place and 4 wins (thanks to somewhat lucky pairing like WAGC 2 years ago: bye 4d- 1d+ (US)6d+ 1d+ 5d-). Here's my report for the BGA: http://britgo.org/news/2018/kpmc. I plan to review my games later, but for now here's a video of my final game where I blew a kill against Czech Radek Nechanicky 5d and a surprising chance to finish in the top 10 (losing early seems to be the way for not so strong players like myself to do this).
And here's a simul game where I worked hard (making a shocking broken shape) to kill Oh Yujin 6p on 2 stones but then missed the final blow (actually Blackie came over and commented how I was trying to kill her which made me lose self confidence, should ignore kibitzers!). Had a nice review with Rob van Zeijst and some strong Taiwanese guy afterwards.

Lost another British Championship. Had a nice opening and first middle-game fight, but game turned when I messed up (more than he just had) in top left corner and had to fight a ko; until played I missed that after c18 atari I was short of libs to push and cut, so shouldn't have j17 tiger, or stop digger deeper hole and don't answer ko threat at j18. And then making a ponnuki worth closer to 1 than 30 points sealed the loss.

I've played a few games on Fox and just got to 7d. Maybe my playing as LZ on another account has helped me learn LZ-style by osmosis: when I play the game I think "what would LeelaZero do?" and it seems to work fairly well in the opening. First game at 7d I was doing well with this strategy, but then forgot about a liberty when about 10 points ahead in yose and game over. Interseting moments/lessons:
- f17 tenuki at r4 is normal (recent Gu Zihao games), but tenuki not bad
- rather than my q4 followup LZ prefers s3
- r8 was me thinking natural q7 was not as good because s9 then better, but LZ says q7 best! (black's move is bad exchange anyway, should just double approach top left)
- d17 I thought maybe f16 was better to settle on top side, LZ agrees by 1% after 5k playouts.
- my fancy h15 "WWLZD" was good
- g15 should h16 push first as he should answer it with h16 connect.
- f13 directly push and cut.
- his k18 big overplay
- my l18 big -35% mistake, as I have ladder I don't need this but can j14 first which means he doesn't have the outside threat 2 capture 3 stones after f15 f14. I did consider that in game, but wanted to be fancy.
- he should f14 firstly though as j14 aji keshi, -20%.
- my e16 empty triangle shouldn't work if he spots g13 counter, so I should g12. But I got away with it.
- initially LZ wouldn't f16 cut but h11 to avoid the squeeze (and then answer black f16 with e18, so sad!), but actually it's not bad and thinks it's better, #1 after 6k.
- because it would break out of squeeze with ugly g12 and atari for ever. f18 slack.
- d6 good WWLZD
- g7 should double hane, I tohught extend good enough but that's because he thank-you pushed instead of sharp j5.
- n5 bad.
- his s16 should r17 as my gobble r11 was good.
- p13 blindly follow middlegame joseki, should just q13 and q11 (or q12).
- by n2 I'm 98%, how can I lose?
- o8 was worry too much about his centre points (he can cut m14). h8 best. I suppose I can let him close centre with a few gotes if I get the big corner yoses in exchange.
- n8 start of losing, forgot p11 lib sente.
- m9/n18 chance to defend q10.
- q10 "oh detritus!" I didn't realise I can atari under and still winning, so went for a terrible ko I can't win. Woops.

Won my first game as a 7d Had the opportunity to play the same 5th line press in that double approach joseki (I think it's good when they have a stone on that side closer than the far corner, in which case the other common choice of hoshi pincer is overplay). Rather than answering he tiger mouthed to defend the other side, and I didn't followup immediately with g4 as g5 push and cut is annoying, but turned trying to get enough extra stones in sente which he allowed (but he shouldn't). This was based on a very similar idea in a bot (FineArt particularly, LZ, Elf too?) joseki I've seen pros trying out too (e.g. http://www.go4go.net/go/games/sgfview/80111 but bots usually g9 extend not g6 atari). LZ agrees with my play and he ended up starting a desperpate ko and I was at 90% by move 40! My first ko threat was good (it's local in that if he ignores the e4 cut doesn't exist) and I picked it over p3 for that reason, but then my answering his threat and letting him win the ko (rather than local threat h1) was just as big a mistake and I went back to even. Also an important lesson with p3 threat, my followup should be q4 to force him to live less efficiently with r3 as his 2 2nd line hane connects of the game give him a comfy life vs if I get p2.

His h7 cut was small and my sacrifice correct, k5 was a nice gift. q17 was WWLZD, if in doubt after approach and knight, attach, it makes opening so easy! n16 better at g17 which I did consider, and LZ would keep the corner not e17. I was doubtful if his top left was fully alive but was too lazy to try to kill it, LZ wants to b17. My n14 was soft, should cap first then jump which is classic technique of slowing down his escape. My r12 wasn't sente, he should ignore and g14 which is a nice move to annoy my thickness and shows why I should have capped, or r12 at h14. k12 cap was then a big mistake in direction which is obvious now with the wall below and how I want to split top from his left group with c10 incoming, not help them join up. I tried killing him but it ended as double ko life (so b15 at b18 to make that corner would have been much better as infinite threats!). n12 was a bluff he believed so I didn't even lose those stones I should have, and then I got a bonus kill at the end.

Great game from my perspective. My first thought though: so many ko's!
Congratulations

Played a tournament last weekend, not so tough opponents but was still interesting trying out AI ideas, playing with my "What would LeelaZero do" hat on:

First game was against this forum's MikeKyle, maybe he will upload our Lizzie review.

2nd game, lower side fighting was the highlight:
- h8 was an overplay, e7 resistance not unexpected and not good but d7 peep one move earlier would have made it easier
- l3 wasn't necessary, just e8 and if he hangs tough with f7 then gentle k4 press was good enough.
- I thought maybe k5 could resist at k7 and it wouldn't be out of Alastair's character, LZ agrees better.
- k6 better at j7 better as I suspected but I wanted to take a lib
- my o4 too obedient, n6! (and then I break out with k7 not l7 hane trying to kill.
- p4 wedge didn't work, game over there but I was pleased to give no chances afterwards with LZ simplify and solidify style.


3rd game was more exciting, featured 2 3-3 invasions and a sneaky ladder; I had some chances to lose in a manic middlegame, will upload later.

Rex quondam, rexque futurus!