Uberdude's journal

[Uberdude]

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?

By this I assume you mean a very-localised search of the peep, rather than this early peep before extending along the side. If I search my slightly old GoGoD for this peep in an otherwise empty quadrant there are only 4 hits (1961, 1964, 2004, 2008). The 1961 one (from Go Seigen) can be dismissed as a non-open position as although that quadrant is empty it is move 100+ and just beyond there are huge walls. ez4u posted about 2 of the others here. So I think it's fair to say AlphaGo's peep is different from those 100 and is behind the current popularity of it. But if you are making the point that the peep locally is not a new move and has a slightly >50% win rate so could that be part of why AlphaGo likes it that's an interesting idea, but I think probably not as 1) I'm still not sure what the AlphaGo initial training set for the policy network was but I think it was more KGS/Tygem games than pro ones, and 2) after those human games there were far more self-play ones so I doubt those 3 extra wins would have much effect.

[Uberdude]

I've still not done much study or tried to get 5d on KGS, but I went to another tournament yesterday (Maidenhead) and won it. Below is game 2 vs Min Yang 4d, lots of interesting fighting and comments in the sgf. I was reasonably please with how I played. The last game was a rather lucky win against Jesse Savo 4d going on 5d from Finland with his endgame blunder and rather less happy with my scrappy play in that game, will post soon. Any comments appreciated.



Final vs Jesse Savo 4d:

[ez4u]

Your theoretical finish in the second game is B+1.5 according to GoEye (so your 64 rather than 62?). At 147 though White missed a chance. White can atari the three stones and Black can not push through the gap between 144 and 146 due to shortage of liberties. This would gain 2 points versus the game. Again (in the theoretical line) when Black plays the atari on Q18, shouldn't White exchange M3 for M2 before answering at Q19?

In the actual game it looks like Black can ignore 182 and cut immediately, winning the fight on the right side. So White does not have time to play 182. If Black gets to block in the corner, I think Black will certainly win.

[Uberdude]

At 147 though White missed a chance. White can atari the three stones and Black can not push through the gap between 144 and 146 due to shortage of liberties. This would gain 2 points versus the game.

Ah yes, I hadn't seen that. Maybe my 145 would be better to simply h14 or h15, but I didn't like the idea of his h12 peep or j12 attach with the bad aji at g10 and only had about 7 seconds per move, but didn't realise it gave him g12 in sente so he could g10 anyway. Luckily it seemed my f10 clamp did actually work to limit the damage, he spent a bit of time there reading out if he could extend and make some mess inside.

Again (in the theoretical line) when Black plays the atari on Q18, shouldn't White exchange M3 for M2 before answering at Q19?

Yes, but I was quickly putting in those exchanges which are black's privilege, so I could play m3 myself there and it should be the same unless he plays the r19 1 point reverse sente next.[/quote]

In the actual game it looks like Black can ignore 182 and cut immediately, winning the fight on the right side. So White does not have time to play 182. If Black gets to block in the corner, I think Black will certainly win.

Hmmm, isn't my corner locally a ko, no time to read semeais! I did answer o18 with s15 though to aim at the cut or at a minimum q12 in sente and he has to connect so I get the 2 stones (but he gets 1 in the middle plus 2 dame so not so amazing).

Thanks for the comments. Did you have any thoughts on the top right fight (e.g. moves 15, 19, 25, 27, 29)?

[Uberdude]

Analysis courtesy of Remi's Crazy-sensei at 4d level, I'll update with my conclusions. But the first most surprising thing is it thinks I was ahead all game: I thought my result on the top right fighting was a disaster! This is only 4d level analysis though, let's see what 6d thinks.

https://www.crazy-sensei.com/?lang=en&g ... ation=game

Here is 6d level analysis: https://www.crazy-sensei.com/?lang=en&g ... ation=game

CS thought black was ahead from the beginning, this may be because I didn't specify the komi in the SGF so maybe it is reading the 0 and thus black leads from the beginning. It shows the black win rate rising at the start of the first fight but then falling back to 50% when Jesse broke through the top right side so with komi maybe it thinks I am losing.

[Uberdude]

Remi confirmed it does read the komi (so I guess no value network which is trained on a specific komi like AlphaGo, or does he do some clever thing to allow changing komi), so, at the risk of being repetitive but also conducting an interesting test, I uploaded a new sgf with the real 6.5 komi on 4d level. This does now better match my feelings during the game that I was leading in the early fight with his funny shape, but then losing once he broke through to the side. CS then thinks I grew a lead with his passive play as I develop the left centre area but then could have lost it if he connected the 2 stones on top around move 115, but as he didn't my lead came back. Then lost it down to even or so with the loss of the 2 stones at the lower left but then cam back again and now that the komi is correct CS agrees that his push (move 183) letting the group die was the biggest mistake of the game (before it didn't highlight that as particularly bad which surprised me), and that my failure to cut one move earlier as ez4u suggested was my biggest mistake. So it's important to set the correct komi for reliable analysis! The analysis is sometimes a bit confusing though because it wants to play some sente sequences earlier than we did: with regards to the endgame relating to life and death on the right side this could be better than what we played, but sometimes I think it is MCTS bad habit of liking to play sente exchanges too early.

4d0komi.GIF (18.51 KiB) Viewed 23257 times

4d6.5komi.GIF (15.5 KiB) Viewed 23257 times

6d0komi.GIF (18.89 KiB) Viewed 23257 times

[Uberdude]

Crazy Sensei 4d7komi (Maidenhead tournament has komi bidding, I can't actually remember exactly what komi this one was) analysis of my 2nd game (I was white): https://www.crazy-sensei.com/?lang=en&g ... ation=game.

ASvsMY.GIF (19.63 KiB) Viewed 23178 times

That big mistake from me then him around move 54 was my s17 seems to be overplay as not truly sente for him to make 2 eyes: seems he can play s15 and win the semeai.

And analysis of my game from the London Open against Kim Seongjin 7d. https://www.crazy-sensei.com/?lang=en&g ... ation=game. I'm interested to see what it thinks of my funky centre jump and the lower right corner joseki. Turns out it actually preferred white (me) pretty much the whole game! It's not a fan of the shimari on move 5, guess it thinks it's too slow. With the ko disaster at the end it thought the game only went back to even, rather than good for black (and didn't find Kim's 2-1 tesuji for better ko).

And vs Haylee: https://www.crazy-sensei.com/?lang=en&g ... ation=game. I forgot to fix the sgf to move the game to mainline though so it's not of the whole game.

[Uberdude]

Last month I had two challenging games in the European Team Championship: against Mateusz Surma of Poland and Pavol Lisy of Slovakia, both European 1 dan pros. I'm feeling a bit disappointed with how they went: in both I played slowly and did fairly okay in the opening and early middlegame but then entered overtime and self-destructed. This was a semi-sub-conscious decision in that I expected to lose so managed my time badly (aka taking time to (over)think about interesting positions I enjoy). Maybe I shouldn't expect to lose: I think I was actually beating Pavol for a fair while when we played last year and I have occasionally scored upset wins against strong players, but it's also not getting my (free) money's worth of a teaching game. Anyway, here are the games with my synopsis plus more detailed move comments and Crazy Sensei 4d analysis:


https://www.crazy-sensei.com/?lang=en&g ... ation=game

Last week I unsurprisingly lost against Mateusz Surma, the new European pro. The opening was reasonable enough, with a big wall and counter-counter-pincer joseki, and to avoid him sacrificing on a small scale I made some attachment, which he resisted, resulting in a trade that I'd seen in pro games before. I got what I thought was a vital point turn in the centre and then pushed out, rather than connecting on dame. He played tenuki to a big point to negate my influence, so I got a nice cover but he plonked a stone in that moyo too. We each took some cash and then I attacked the weak group, but only chopped off a small part and my wall, that never connected on dame, suffered. As I was behind and short of time I played a few moves that didn't work and resigned.

https://www.crazy-sensei.com/?lang=en&l ... 106&move=0

I lost against Pavol Lisy 1p. In the opening he picked an unusual move to grab territory and I got a big wall for a side moyo, which he then wedged (not in the place that has a miai two-space extension out of respect for the wall, an advanced technique I'd studied recently). But, rather than settling on the side, he attached on top of my pincer and then pressured my corner, and I moved out with a shoulder hit. I then spent a long time deciding whether to k16 jump on top or f11 surround his group, and chose the later, but maybe it was a mistake as only sente for a seki, not a kill, due to a sente move of his. So he raced ahead on territory, I made some small weakish groups in his Chinese opening side, but he played rather small yet aggressive moves denying a life in his final corner and I didn't really know what to do about my centre potential. By now I was in overtime and trying to get as many moves down as possible, so, when he played a sente move on one of my groups, I played some extra moves which ended up filling in my own liberties, thus I killed myself and resigned. That was rather unfortunate as I think that previous seki had now become a possible sneaky kill.

Detail:
- normal move is take, but I have seen this one before. Result seems ~fair. But maybe Pavol wanted to take cash because I won't be as good at using influence.
- this middle-game joseki respects the wall by not wedging where you have miai 2-space extensions.
- pincer this side or above? I think this one is more common (even though making territory with wall) due to b cramped. Worth noting though that often with this wedge and a big wall below black has the corner (e.g. if black captured instead of connect and then 3-3 joseki) so this extension also serves as making eyes for the wall, but here the lower left corner is open so is available for potential eyes for my wall. Thus maybe c14 more plausible.
- normally black 23 is first attach under at b10 as a probe: normally white hanes behind, but if white cuts and b12 push to resist then black jumps to c14 and white ends up wanting to help the corner and b9, so doing this first is maybe a mistake? I answered as normal planning to then b11 and b12 push if he plays b10 next trying to turn it into a mistake, but I didn't expect his attach on top which was presumably him realising his play out of order means I wouldn't revert to joseki.
- agressive!
- big think trying to work out how to get out/settle with good shape. Also considered c14 trying to get b11 peep but read it didn't work (not sente, he can c15 and cut, or he can c9 sacrifice to get b11 sente himself). f16 g16 g17 was another idea trying to sacrifice to get sente strength and then attack side group, but makes him strong and lame attack.
- I can't capture at b14, but I did wonder about a sacrifice and then e12 (if f10 then f11) try to kill below and chase into wall but seemed impossible.
- relieved he didn't push and cut.
- Big decision. Here or k16. I can't remember if I thought/hoped this was sente, but it seems to only be sente for a seki because black c15 is sente. So probably wrong. But CS likes it.
- honte, but is it too slow given b's lead on territory and my influence? But if I q5 now black can play here to attack me and q5 loses and attacking meaning on r9.
- try to catch up points. Treat as miai with other corner thanks to honte before.
- maybe I should kosumi to live in corner?
- shcizophrenic, in overtime now and my game loses balance.
- w100: this was a get one of 15 moves in 1 minute down move, but I think it's actually sente to kill left as c15 no longer sente.
- w104: retarded, just live at L19 you dumbass!

[Uberdude]

I've not really done much study yet, so this month I joined BIBA's new online league (http://www.bibabaduk.net/index.php?p=pLeague). I'm at the bottom of the top group of 8 (not so tough opponents in Inseong's) so should get good games, plus the reviews/lectures/problems/teaching games. This should be good preparation for the British Open in early April, and British Candidates in early May (but I might be away for that). Also there's the Cambridge tournament this weekend. Now to work through these rather difficult 24 problems by Monday!

[dhu163]

the online school seems cool. The left vs Pavol is unconditional life, B can give up the two stones

[Uberdude]

So I'm coming to the end of the BIBA online study month. I've played 6 games (4 losses and 2 wins, including a surprising one against Benjamin D-G French 6d) and even did some of the homework problems (will try to finish them next month). The reviews from Blackie were good, but the teaching game was rather a missed opportunity as I suicided early.

I also played another European team game, which I ended up winning by half a point! Here's my synopsis:

My game was against 4 dan Armel-David Wolff, who I'd met at the World Amateur last year. I thought the opening was slightly good for me once he didn't go for complications by pincering my bottom left approach and I invaded his dubious shape on the left side. Inspired by AlphaGo, I went for the kick and jump when he approached the top left, aiming to overconcentrate him on the side but he extended further on the 3rd line so I didn't get a chance to play the 5th line shoulder hit and he snuck in the corner and lived (for now) in sente which was a bit annoying. I took a big thick endgame move and grew my lower moyo, and he reduced it by aiming at an invasion. I answered in a thin but good endgame way, but maybe I should have ignored and split his right side (splitting groups was big and created powerful follow-ups) which he then fixed. He resisted my attempt to get a double sente endgame and attacked my left group: deciding discretion was the better part of valour I lived submissively. By now I was in overtime and lost of a lot of points when he broke into my lower side (I missed a tesuji to kill him due to his shortage of liberties from a bad push) and the game got uncomfortably close. There were a few minor ups and downs in the endgame but in the end I fortunately managed to squeeze out a half point win (after losing the final ko).

And Crazy Sensei analysis (it thought I did actually fall behind in middle-game which seems plausible, and then caught up in endgame): https://www.crazy-sensei.com/?lang=en&g ... ation=game

[Uberdude]

So my goals are:
- at least play in if not win British Championship title match next year
- get 2400 rating
- get solid 5d KGS from non-blitz

Last weekend was the first stage of the British Championship and (having got permission from the boss to play) qualified for the next stage (8 player round robin Challengers' league) with 5 wins out of 6 which sounds good, but actually I played rather poorly and had easy opponents (all wins against people > 2 stones weaker) so am not particularly confident about the first goal above. I didn't beat any of the other 7 qualifiers (dhu did well and won the tournament with 6 wins and beat 5 of them, maybe he'll post soon) and lost to Andrew Kay (tried to have fun in the opening and lost early) so actually lost rating points despite 5/6 down to 2363 (meaning dhu overtakes me), so 2400 is a way off. Interestingly enough the 7 qualifiers are exactly the top 7 entrants by rating (excluding Walter Gregory 2d who hasn't played in 20 years so his grade is probably too high (though he did give Andrew Kay a decent fight) over Tim Hunt who was 10 points lower). And I've not played on KGS. So a rather unsatisfactory warm up for the Challengers' league, I'll need to play a lot better to do well in it.

P.S. one of my wins was against Sam Bithell 1k, a relatively new player from the Milton Keynes and Cambridge clubs. I had beaten him on 4 stones at the club recently, but failed to clinically kill him and he did well with 3 wins: it's for good experiences like this it's nice to let 1ks play in the Candidates.

[Uberdude]

My 2nd game from Candidates (because it was broadcast and some people having trouble with sgf):



Game 3 against Andrew Kay, the only strong player I played in the tournament. Funnily enough also ended by resign at move 107, but really it was lost already at move 20. My decision making sucks.

[dhu163]

I added a few comments. For example, there were many better ways in kay's game, and he played overly aggressively.

[wolfking]

Game 3 against Andrew Kay, the only strong player I played in the tournament.

Andrew, in your game against Andrew Kay, if white play at N18, seems to me it's (W)eye vs (B)no eye and white at least get a favorable ko out of it?