LeelaZero adventures on Fox

[Uberdude]

Played a game with the new Elfv2, it didn't really get a chance to demonstrate its strength as the 9d human opponent made a order mistake in that common joseki mentioned elsewhere was down to 5% by move 55. Still it was interesting to see how leisurely Elf was in the subsequent attack, happy to give a load of territory and push along the 3rd and then 4th line in gote on the left, allow the weak black group to jump out, but thinly, and only play sharp moves later. In fact the j3 push in was my choice of a move with few playouts but higher winrate (and I checked it remained so playing out variation) than the passive g2 that was Elf's highest playouts move because I wanted to make him resign (he did so at 118) so I could do something else.

P.S. Elf thinks it's a big mistake (-13%) for white to play a 2nd 4-4 if black plays a 4-4 and a 3-4 (slightly more if 3-4 facing instead of skew).

[Uberdude]

Another Efv2 game. Highlights were:
- how big a mistake it thought black's atari of 19 to live inside was, a result I supppose many humans (myself included) would think was reasonably even given black lives in sente and can pull out the q12 stone. See var for suggestion in this sort of middlegame joseki shape, a nice lesson (I've seen same sequence suggested by bots before).
- t11 (bad) vs s11 (correct) connection makes a difference to a semeai if white wants to get s6 as kikashi, Elf so sharp to instantly see this.
- m10 to stop the wall getting attacked at black n10.
- j4 should be o6 to prevent white easily connecting 2 groups. Topology is king.
- move 100 was a misclick of atari, but still winning and soon back to 98%.

[Bill Spight]

Another Efv2 game. Highlights were:
- how big a mistake it thought black's atari of 19 to live inside was, a result I supppose many humans (myself included) would think was reasonably even given black lives in sente and can pull out the q12 stone. See var for suggestion in this sort of middlegame joseki shape, a nice lesson (I've seen same sequence suggested by bots before).

It appears in Sakata's Tesuji and Anti-Suji of Go.

[Uberdude]

Another Elfv2 game with some blindspots. Elf thought j6 was a big mistake, and it all went as expected for a while. However, when white played h4 atari the winrate dropped a lot: black can connect and is safe against e3 with g2 and g1 tesuji, but did it misevaluate the ladder? Anyway, things were looking good again in the top left fighting, thne f12 was a semi-blind spot (it was a move it explored before but winrate for white improved when played), but then it didn't expect white to be able to connect at c14 so f14 etc weren't good moves. In the end it could capture white but got horribly squeezed (and played some silly sentes to postpone doom).

[Bill Spight]

I wonder, are we talking about specific blind spots, or some kind of horizon effect? The neural net bots do whole board reading, and unless they recognize certain patterns they will not necessarily explore a local region thoroughly.

I have noticed DDK errors with Deep Leela, particularly as the dame get filled. Some of them may be explained by hoping for the opponent to make a mistake, but not all. It makes me think that if I ever played Leela 11, I should never resign.

[Uberdude]

Another Elfv2 game (white), I think the opponent was a strong human 9d. First talking of blindspots, there's one on move 2. If Elf plays 4-4 as black first move then it's at 48% it only expects white 4-4s in reply. But if white plays an adjacent 3-4 then Elf is surprised by this innovative move and black winrate drops by 5%! Anyyway, this game was high quality with Elf gradually taking a lead but some ups and downs. Then in the endgame with Elf leading for move 188 obvious sente b2 and e11 atari were neck and neck, so I picked the #1 at end of byo yomi e11 atari. It was a massive mistake as black has e1 tesuji, which was not mainline after e11 atari but was spotted once played. Even then it still thought f14 atari was 80%, but then when black collapsed the winrate plummeted. Strangely replaying it now I get a different story. Anyway, black returned the favour with 227 blunder (should s2) because Elf now started a ko and had one good threat, whereas black had lost his h6 atari threat. The 2nd game variation is Elf lsoing to a minigo bot on super GPUs.

elfv2 double reversal.PNG (188.15 KiB) Viewed 20351 times

[Uberdude]

By chance I managed to play the same guy (some chat said AI, some human) as the last Elfv2 one, so I tried LZ 205 for comparison (White again). It was a really nice game, and I'd say won in a more convincing fashion and no blindspots.
- the first fight featured running with a knight's move (a haengma now on my radar) and ended with a dramatic trade LZ thought favoured white. Black looks pretty nice too, but bear in mind white got a ponnuki whilst the dead white stones are still on the board so have aji, as dramatically demonstrated later.
- h9 jump was prevent white h10 nose attachment, n7/j3 other ideas of LZ.
- b16 invasion / 2nd line pincer was fun (LZ's choice too).
- n8 bamboo wrong timing (k7 better), as lost a lib for w's throw in ko.
- sabaki masterclass to demonstrate black's threat couldn't kill, note the many diagonal moves to build eyeshape and use of dead stones to counterattack and live.



For the winrate graph bear in mind Elfv2 has more extreme opinions, so putting this game into Elf it thinks white did better than Elf did itself in its game.

lz205 vs elf reversal guy.PNG (150.25 KiB) Viewed 20273 times

So although some other big tests (at lower playouts I expect than here) showed Elfv2 as stronger than LZ head-to-head, against Fox 9d this little sample suggests the reverse to me.

[Uberdude]

Another LZ 205 game, against someone who turned out to be a 6d Chinese student studying in England! LZ thought o5 extend was a mistake and a good trade for black getting up to 72%. But then the winrate declined to even again with just a few unexpectedly good moves from him (q15 being unexpected was weird, an obvious option!). But then b14 cut tesuji was a mistake as LZ sensed a weakness of white's centre group, and managed to kill it when he declined several chances to compromise.

hill kill.PNG (194.09 KiB) Viewed 20198 times

[Uberdude]

This one was interesting, LZ 205 black against a guy running AQ on his oneplus5 phone. I'd lost to him before (by 0.5 not sure using which bot), was surprised that a phone could do better than desktop with GPU as I was losing most of the game until a reversal in late endgame, was there a mistake or komi problem? Specs of phone are CPU: Octa-core (4x2.45 GHz Kryo & 4x1.9 GHz Kryo), GPU: Adreno 540, I suppose AQ can make good use of the phone having 8 CPU cores?

vs AQ.PNG (157.32 KiB) Viewed 20041 times

[Bill Spight]

This one was interesting, LZ 205 black against a guy running AQ on his oneplus5 phone. I'd lost to him before (by 0.5 not sure using which bot), was surprised that a phone could do better than desktop with GPU as I was losing most of the game until a reversal in late endgame, was there a mistake or komi problem? Specs of phone are CPU: Octa-core (4x2.45 GHz Kryo & 4x1.9 GHz Kryo), GPU: Adreno 540, I suppose AQ can make good use of the phone having 8 CPU cores?

vs AQ.PNG

Looks like an evaluation problem to me.

[Bill Spight]

Here's why I think so.

White 314 loses 6% ????? Well, Black is considered to be ahead at this point, so that may just be the win rate approaching 100%. But that underscores the problem of using consecutive win rate estimates to indicate mistakes.

White 310 loses more than 3% ????? Here there is no excuse that Black is ahead. even after White 310 Black is estimated to be ahead.

White 302 gains 1/3 pt. (I am using territory scoring.) Is this a mistake at area scoring? Possibly, depending upon the ko fight. At territory scoring canonical play is for White to take the 1/2 pt. ambiguous sente on the top (White 304) and then take the 1/2 pt. gote at 303. Then Black can take the ko at 302, but White has a local ko threat. (Edit: I have looked at the ko fight. Black is komaster, but I don't think he has enough ko threats to make White filling the ko superior.)

Black 301 gains 1/2 pt. It is miai with 303, but worse for ko threats. Canonical play is at 304, ambiguous 1/2 pt. reverse sente.

White 300 gains 1/2 pt. Again, canonical play is at 304.

White 298 gains 2/3 pt. (ambiguous reverse sente). Canonical play is at 299, gaining 3/4 pt.

White 294 (saving 2 stones) loses more than 1%, with White ahead?

Black 267 gains 3%????? With White ahead? Surely that indicates a previous misestimation.

White 256 loses 6%????? Even versus the variation? Really?

[Uberdude]

LZ has improved (in self-play Leelo at least) quite rapidly in the last few weeks, here's a game (LZ black) with 214. Opponent held his own fairly respectably until he died in the wrong way at top right (because he didn't make miai of saving the dead group or doing mean things to bottom right corner). A fun ko at top left, with him making nice use of the technique of playing out lower side to create some large ko threats, with LZ's decision to end the ko because it knew it could sacrifice the lower stones in an effective way to build some centre points was impressive (answer threat was not bad, but just something like 96 instead of 97%).

[Uberdude]

A game with latest netwoek #223. Not sure if the opponent was a bot, he always used 25 seconds per move for a while and only lost advantage slowly, but then messed up a fight and didn't resign for ages. I thought LZ's judgement of when to play solidly to settle, when to sacrifice, when to play severely was interesting. LZ's winrate slowly grew until he messed up a fight and game over. Highlights:
- move 26 white should have d9 sente first to stop black making some eyespace there.
- 41: LZ wanted r7 extend but lighter r8 was a promising blue circle, I played out a few moves and winrate stayed higher so I picked it.
- 48 was mistake, should n4 or l7 and make a mess. So why didn't LZ want to atari there before j4? And why did it want to atari there now rather than directly l7 (-1%). I wonder if something to do with white j2 in a sequence wanting to be gote. Or is it just if black makes the territory it's double sente yose?
- 65 don't often see LZ make this block, but quick life and makes r11 wrong
- 69 cut was preferred to (natural to me) o16 (-7%) getting out even though it then didn't run with it and just went back to living in corner. I think the option of q12 atari to help reduce a w moyo and make w bad shape was why.
- 76 o16 slow, should f17
- 77 interesting that LZ picked this way to attack, it is mutual base of 2 unsettled groups but struck me as slow compared to top side (e.g. g17 -3%). But it judges strength and weakness of groups here as urgent and doesn't mind if white extends on top, because by attacking this group we can expect to keep top under control
- 81 voila, black d9 group is now essentially alive whilst white group weak and only now does black play top side in good rhythm.
- 83 another lesson, attack from 2nd line not centre!
- 95 W should answer this peep in centre and black is 65% (see sgf var - it's in a way that means if black cuts white group has exit route and black has ugly shape. e18 was an invasion point but he got timing wrong. Now he starts making big mistakes, maybe his bot crashed.

223 vs bot never resign.PNG (111.19 KiB) Viewed 14178 times

[Uberdude]

Down to 8d, many bots at 9d on Fox. This was clearly against a human 8d, a beautifully clear example of the LZ strategy of taking territory, give walls, make a live group inside the moyo and 99% by move 80. Of course the 8d opponent helped with his slight mistakes, but still impressive. The paratrooper of move 50, kind of on the sector line aiming at p11 or q11 for sabaki was a highlight.

[Uberdude]

LZ 229 (black) using a ko threat factory (main branch):