LeelaZero adventures on Fox

Uberdude · **#61**

Played a game as Elf v1 against another person using a bot, an Elf v1 v0 mix on 1050Ti. There was an interesting moment at move 65 where my Elf thought it was doing well and winrate spiked up to >60% but then quickly fell because it had missed a vital ladder that meant it couldn't blockade and kill the white group (see variation on move 71 for ladder). Later a big group of mine died.

Uberdude · **#62**

Although it's been interesting playing as LZ on Fox, I've been unsuccessful in my aim of seeing how strong LZ is on my moderate hardware vs pros. I don't recall playing any confirmed pros to date, and most opponents at 9d are probably bots too. There may be a selection bias here in that many humans don't want to play me with my obvious bot name, but I've also heard the pros online tend to know each others' username and only play each other. So there could be quite isolated pools of players. Also I've seen some pro accounts (wonfun = Won Seongjin I think) with not particularly great win ratios, whilst most of the bots I've played have good ratios so whilst this could mean the bot accounts are stronger than a top pro like Won if they don't play each other you can't really say.

Anyway, today I played a game against Chinese chenyuch with a p badge so I think that means a confirmed pro (I couldn't find a pro called Chen Yu Ch??? though) with 125 W 157L. LZ beat him easily (resign at 146): a steady decline in winrate where probably more than half of his moves which weren't obvious to me as a mid dan amateur were viewed by LZ 191 as a mistake. Biggest mistake was in the first fight at top right: although he killed the cutting stones LZ got a nice force on the outside at n13 (with later squeeze very handy in another fight, in fact LZ thinks m14 good shape defence might be better at o14 to avoid this) and could still make a gote life in the corner using the aji (o18 should be q17 to avoid this). Once the 20s byo-yomi started LZ also played a few moves which with more time it realised were slight mistakes (but 90.5% is not much worse than 90.8) though I did do a little of promising blue circle promotion. Mainline is the Fox game, he resigned at 146, the variations are me following Ke Jie vs Chen Yaoye and Park Junghwan vs Mi Yuting in Chinese league.

Attachment:

chenyuch.PNG [ 148.31 KiB | Viewed 1618 times ]

Uberdude · **#63**

Checking the user info of chenyuch I see it says he is 2p with a name 季建宇 = Ji Jianyu but I can't find that name in rating lists.

John Fairbairn · **#64**

It looks like you may be misreading the name. At least there is a 李建宇 Li Jianyu 2-dan, from Cangzhou, born 1990. He was a late (under-25) qualifier as a pro and is not very active - I only have a couple of his games.

Uberdude · **#65**

Thanks John, I was using an online OCR tool plus google translate! I see Li Jianyu in https://senseis.xmp.net/?ChineseProRatings20130430 at #238 of 320 with 2230, with Zhao Baolong (of EGF fame) at 175 with 2277 and the top 10 at 2600+.

Uberdude · **#66**

Played a game with the new #195 network. It was the first game in a long time that LZ had a ladder delusion (fast game at 15s a move so averaging about 3k playouts, needs 11k to realise ladder is bad and not play n10 atari but p12 better). LZ thought it recovered, with the lead changing hands a few times (though as white on Fox has 1 less komi than LZ assumes) but lost by 1.5 in the end.

Attachment:

195 ladder delusion.PNG [ 227.48 KiB | Viewed 1386 times ]

Uberdude · **#67**

Someone played mirror go as white against me for the first time. LZ (I used the super trained 157 network 9229) handled it ok in the end, making a centre fight in which tengen meant it won a semeai. I didn't help it (e.g. I looked at mainlines and checked to see if it was tring to make double ladder and wouldn't have stopped it if it was going to do something dumb), though I did give it 2 byoyomis when then winrate briefly dropped below 50% around move 70: its instinct was to capture the top white stones, but then white's komi is helpful, after 25 seconds or so it decided saving the stones was better which makes sense as then sente/tengen becomes valuable.

ez4u · **#68**

If you look at the analysis, was LZ looking at White continuing to mirror Black or was it calculating based on White choosing other "better" alternatives on each play?

Uberdude · **#69**

Sometimes LZ expected the mirror move, but generally not. I noticed that when there were 2 hot areas of the board (as often with mirroring) it tended to expect play to continue in the area of the previous move. I did sometimes play his expected mirror move before he played it to give LZ more than the 15 second overtime to think.

Bill Spight · **#70**

Uberdude wrote:

Sometimes LZ expected the mirror move, but generally not. I noticed that when there were 2 hot areas of the board (as often with mirroring) it tended to expect play to continue in the area of the previous move.

Mirror go is an effective strategy if the temperature drops below 2 x komi, something that tends to happen as the board fills up. So the player being mirrored needs to keep the game exciting. If LZ's last play raised the local temperature, then it would make sense for the opponent to reply locally. Perhaps that is what was happening.

Quote:

I did sometimes play his expected mirror move before he played it to give LZ more than the 15 second overtime to think.

Naughty, naughty! :blackeye:

lightvector · **#71**

I've always had it in the back of my todo list to test a little hack, where if the opponent has been mirroring for a while, adjust the policy net to put, say, an extra 20% of its probability mass on the mirroring move for the opponent in addition what it would have normally (taking that mass equally from all moves). After all, supposedly the policy net is as one of its functions supposed to be predicting what move will be played, and predicting that they will mirror has got to be a pretty good prediction...

Or more mechanically, it should make the MCTS search prefer to go a little deeper down lines where the opponent continues to mirror, and to make it so that the average winrates at interior nodes in the tree are biased just a little towards the winrates that would happen if opponent will play the mirror response rather than what the policy net would ordinarily predict. In particular, this should make the search "try" to slightly favor lines where it thinks the opponent continuing to mirror is extra-bad for the opponent, and also to more carefully read those lines.

Will get around to it at some point, once my experimental bot is more developed in other more-important ways...

Uberdude · **#72**

A brief interlude for a LeelaZero adventure on KGS. I played with the super 157 15-block network 9229 against latest 40-block #197 used by petgo3 account on KGS. I have a i3-8100 with 1060 3GB GPU, it has "i7-7700HQ, Nvidea GTX 1060" so pretty similar, time setting was 5m + 10x20s, so this was a battle of a stronger network with fewer playouts vs a weaker network with more playouts at time parity and pretty much hardware parity too. My LZ won by resign, despite me making a -7% misclick (h10 should g10) in a middlegame fight. petgo's game-deciding blunder was m18 atari which makes c18 work as k17 is now black atari so f17 pull-out works, I guess it just didn't have enough playouts to see that. I was getting really nice playout numbers with this shallower half-precision network, often 30k+ when expected sequences happened whereas I get just ~3k when using a 40-block network at similar time settings.

Attachment:

vs petgo.PNG [ 140.19 KiB | Viewed 1015 times ]

splee99 · **#73**

Note that Petgo3 sometimes uses as low as 1k playouts per move (randomly sampled at KGS), while at move 80, the 157a used 120k playouts per move. So this win of 157a should be expected.

Uberdude · **#74**

Had a strange interaction with some Chinese guy today, probably using a bot. When he asked what GPU I used and I replied 1060 he accused me of lying and was very insistent on it. I presume he couldn't believe a 1060 rather than a stronger GPU like a 1080 could be 9d or have my win stats (I lost the game, so it's not like I beat his stronger GPU with my weaker). Here's the Chinese chat in case google translate lost something. Weird, is it so surprising a 1060 can be 9d? I didn't surprise me.

Uberdude · **#75**

This game was interesting, demonstrated the importance of getting the timing of a peep wrong. h13 peep would have been a good at move 57, but he played it at move 59 after white had defended with the f10 kosumi and it lost 15%. This kind of kosumi after a knight move jump is a good solid shape to learn and I thought this was defending the g12 weakness whilst reaching out to g7 (with bonus of d10 followup into left side), but it also means the h13 peep becomes resistable with j12 attachment as then it ends up perfectly played to defend the g10 cut when you play the obvious sequence trying to break out (see variation).

sorin · **#76**

Uberdude wrote:

This game was interesting, demonstrated the importance of getting the timing of a peep wrong.

What are the variations in the SGFs from your online games using LZ - are they auto-generated by Lizzie somehow (if so, how to do that)? Or did you add the variations manually, and then ran Lizzie in auto-analyze mode?

Uberdude · **#77**

sorin wrote:

What are the variations in the SGFs from your online games using LZ - are they auto-generated by Lizzie somehow (if so, how to do that)? Or did you add the variations manually, and then ran Lizzie in auto-analyze mode?

I made them manually with pondering on, choosing top candidates moves of LZ. Some are me exploring the promising blue outline circle candidates, i.e. a move with lower playouts than LZ's filled blue first choice with most playouts, but with a higher winrate; so I will play that move and allow LZ to devote further playouts to it and if the winrate stays high once it has a similar number of playouts then I'll play it (e.g. l12 atari at move 88 leading to the trade in preference to initial choice of peaceful j13); sometimes the winrate will drop, typically because LZ then finds a strong counter to the move for the opponent that makes it not as good as it first thought (e.g h15 at move 34, or e8 at move 32: that had higher winrate than k17 with a few playouts but that's only because its instinct was that black would answer with nobi and then it was a good exchange, but it soon decided black would hane in response and then it wasn't a good exchange). So this is me manually increasing LZ's MCTS exploration factor so that it can find moves in a 30 second (or whatever) byo yomi that it would naturally find as best only with more time. When I usurp Lizzie's blue circle I think I do usually pick a move which is a stable better move, though occasionally I can hurriedly pick a weaker move which further analysis reveals as not as good as the original choice. Other times it's just me satisfying my curiosity of "what would LZ do if this happened?": I'm not just a dumb move-relaying monkey but a go player who uses this as a learning opportunity (for those unaware, shift and left/right arrow in Lizzie allows re-ordering of the game tree which is very useful to move the game line to lizzie's mainline to keep the sgf tidy and winrate graph synced). h13 peep on move 57 was me after the game exploring "Is this a good time to peep?". Answer is yes, LZ's recommended move for black at that point was l16 rather than n18 but it also thinks h13 is good as seen from the "(0.4%)" annotation, meaning after that move that player's winrate increased by 0.4% which is more than the 0.2 of n18 block (it's not unusual for winrate to have a few-tenths of % adjustment after move played vs before), and the variation shows LZ's expected continuation.

sorin · **#78**

Uberdude wrote:

(for those unaware, shift and left/right arrow in Lizzie allows re-ordering of the game tree which is very useful to move the game line to lizzie's mainline to keep the sgf tidy and winrate graph synced)

Thanks for sharing your methods. You are quite a power-user of Lizzie - I had no idea about this feature!
I wish Lizzie had an option to automatically create a variation for the top winrate move when that differs from the current move.
The current way, it seems that the only option is to annotate each move with a +/- percentage, but that doesn't tell what the better move is (which would be very nice to know when there is a big %-age drop for instance) - one has to reload the game in Lizzie to find that out.

Thanks also for sharing your adventures with LZ - it is very interesting!

Uberdude · **#79**

This super-157 is a plucky little fighter: it managed to hold its own running on my 1060 and even increase winrate as black from 47 to 55% * in middlegame against another bot player using the more-than-4-times as powerful dual 1080Ti (and I presume a 40-block network) in a 15 second game before one of my usurping blue circle decisions (t11 over s11, it didn't expect white t8) turned out to be bad (but maybe it would have lost anyway). There was a big semeai that ended in seki.

* However, this seems to be a misjudgement of the position, at move 204 LZ #198 thinks black is down at 28% already. I think the smaller block networks do have more trouble with semeais involving big groups. Todo: would be interesting to plot winrate graphs from various networks for this game.

Uberdude · **#80**

So LZ 157a and #198 have quite different opinions about who was winning that game in the middlegame. First graph is 157a on 10k playouts, 2nd is 198 on 1k (which is a bit lower than equal time) with extra playouts at a few key moments for greater accuracy. Basically 157a thought the fight was pretty even and even sometimes good for it, 198 says white always good. I believe 198 more. It's also weird that after the seki is clearly formed 157a still wanted to keep playing mutual dame in the seki, did it not understand it was seki (such moves having equal winrate to other moves better for points elsewhere I can understand, but better point moves having winrate drops is weird, maybe it thinks only chance to win is if it plays self-atari it the seki and opp ignores? Or is the seki just too big for 15 blocks?). And third is elfv1 on 2k playouts. Elf tends to like black more than LZ as we see here to start with, but once the big fight gets well and truly underway it also thinks white takes the lead before 157a realised it was in trouble but still thinks there's quite some swings to almost even.

Attachment:

big seki 157a 10k winrate.PNG [ 193.62 KiB | Viewed 461 times ]

Attachment:

big seki 198 1k winrate.PNG [ 226.13 KiB | Viewed 461 times ]

Attachment:

big seki elfv1 2k winrate.PNG [ 205.12 KiB | Viewed 460 times ]

P.S. Yeah I should export these winrates as numbers to graph them all together.

LeelaZero adventures on Fox

Who is online