LeelaZero adventures on Fox

[Kirby]

I don't quite follow it - black plays q4, and r4 is already there, then. Which move number?

[Bill Spight]

I don't quite follow it - black plays q4, and r4 is already there, then. Which move number?

I think he means Q-03 was a misclick for R-03, at move .

[Bill Spight]

Well, that was weird: just had my first loss, and LZ thought it was winning the whole game but lost according to Fox by 17 points, and that was a no komi game as black against an 8d. It lost winrate from the lower side fighting, but then thought it recovered. Counting the game independently with CGoban and myself I get a black win by 17, a Fox bug? At the end of the game I clicked the Count button and then a dialog for automaric counting came back saying "W+17 'leelazero7'" so the colour of winner was wrong but name right. I, perhaps stupidly, clicked ok because another game I won on time displayed the wrong B/W colour so I assumed it was a similar UI glitch and wouldn't actually make the game result incorrect. Is it possible in Chinese client for a player to change the winner field?That's annoying to lose the chance to double promote to 9d.

loss winrate.PNG

Weird.

BTW, what does LZ think of G-15? Human go proverb: Don't let the double keima get away.

Edit: B109 surprised me.

[Uberdude]

I don't quite follow it - black plays q4, and r4 is already there, then. Which move number?

I think he means Q-03 was a misclick for R-03, at move .

Indeed.

BTW, what does LZ think of G-15? Human go proverb: Don't let the double keima get away.

Yup, LZ thinks he should have played there (but not a big mistake, <1%, not defending at d12 next was a little bigger at -1.5%) and was quick to see it as the vital point and more important than answering approach to 4-4 (a common lesson from AI reinforced to me doing these games, answering approach is not so urgent when there are important moves elsewhere).

Edit: B109 surprised me.

Me too, I was going to highlight as one of the moves of the game (along with him finding the d2 tesuji, LZ saw it but I thought he might not). When LZ played the atari beforehand it was planning to just L12 extend which is a shape in my vocabulary, but then it found and liked that elephant jump as #1 by the end of the byoyomi (but it was at 95% by now so many moves all over the place were good enough). I liked how LZ kept tenukiing from the left group because it wasn't actually killable.

[Uberdude]

That 8d who beat me was stronger than I expected a Fox 8d to be (much stronger than the 7ds so far). This game vs 8d is more like what I expected, same ladder mistake in that joseki, but quickly recovers and then crushes. 26 and 32 were the highlight for me: 26 doesn't actually capture the stones like the net would, but the net gives black good forcing moves so you want to capture it more efficiently or let black run with a heavy group, and 32 is a nice "take care of cutting stones aji whilst pressuring the black group" move, LZ is good at finding these alien-to-me shapes quickly.

8d reversal winrate.PNG (194.6 KiB) Viewed 26349 times

I think I'll keep using 157 for a bit as it usually beats 8ds with no trouble and we have the following datapoints:
- played non-working ladder atari, aborts, recovers a bit, I misclick -10%, can't reverse game and loses
- played non-working ladder atari, aborts, recovers a lot and crushes
I wonder if it can still beat a strong 8d (cenovis92 account was Korean with about 3:1 win ratio) after ladder mistake if I don't misclick.

Here's another win at 8d, was a nice smooth game with LZ taking territory and then playing a well-balanced reduction, with a fight ending in seki later.

[Uberdude]

Little test of #188 on a Macbook Pro (CPU only version of LZ), was a 15s game so that's only about 50 playouts per move if it's unexpected, got up to a few hundred when it can reuse tree with expected moves. Lost due to creating a ladder which didn't work (it wanted to play the failed atari but I played another choice tenuki to see if there was any chance to get winrate climbing up but there wasn't so resigned at about 6%. Opponent was a 65 win, 2 loss player who was also a computer as I found out via google translate chatting in Chinese after



Game chat:

leelazero7 [8D]: Hi, I am LeelaZero #188 on macbokpro
商业围棋 [8D]: 苹果电脑？= Apple computer?
leelazero7 [8D]: 是的，没有GPU所以无法读取梯形图 = Yes, there is no GPU so the ladder cannot be read.
商业围棋 [8D]: 你输多了 = You lost more
商业围棋 [8D]: 投降吧 = Surrender
‘商业围棋’ W+ Res ‘leelazero7’
leelazero7 [8D]: Thanks
商业围棋 [8D]: 没有gpu打不上9d吧 = No gpu can't hit 9d.
leelazero7 [8D]: 你是人还是电脑？ = Are you human or computer?
商业围棋 [8D]: 我也有电脑哈哈 = I also have a computer haha.

[hitlab]

Well, that was weird: just had my first loss, and LZ thought it was winning the whole game but lost according to Fox by 17 points, and that was a no komi game as black against an 8d. It lost winrate from the lower side fighting, but then thought it recovered. Counting the game independently with CGoban and myself I get a black win by 17, a Fox bug? At the end of the game I clicked the Count button and then a dialog for automaric counting came back saying "W+17 'leelazero7'" so the colour of winner was wrong but name right. I, perhaps stupidly, clicked ok because another game I won on time displayed the wrong B/W colour so I assumed it was a similar UI glitch and wouldn't actually make the game result incorrect. Is it possible in Chinese client for a player to change the winner field?That's annoying to lose the chance to double promote to 9d.

loss winrate.PNG

This was a game using a special rule set of captures counting double the points.

If you look down your game list, there is a column that says "none". This game has "一子千金", meaning this special rule set applied to this game.

If you don't want to play these type of games anymore, careful not to accept auto matches with the word "一子千金" at the top of the box

[Uberdude]

Thanks hitlab! I was aware there was that new mode, but didn't realise you could get automatch challenges from it if you never clicked it yourself. I'll look out for that word in automatches in future.

[Uberdude]

Another #188 with low playouts (bit more this time as 30s byo), easy win vs 7d (luckily the ladder on opening was good for white!) and even just managed at end of byo-yomi period to find the shortage of liberties tesuji with move 84 as superior to capture (but both >97%). And then a nice kill at top side with another of shortage of libs, a classic tsumego that I probably wouldn't spot unless you told me it was a tsumego problem but 188 spots with just a dozen playouts, so quick to see the vital shape points.

[mitsun]

The life and death problem on top was nice, but perhaps not classic. B123 was suicide, pretty much forcing W to take the killing liberty B could play F18 instead to live.

[Uberdude]

Nicely spotted mitsun! Analysing now on my more powerful PC LZ 188 does find the f18 atari after about 400 playouts, but it still doesn't become the #1 playouts move after 30k. It does have a higher winrate of 2.7% vs 2.3%, but seeing as everything loses it doesn't really matter. I mean more a classic technique of tsumego of creating the shortage of liberties to keep the eye: it's something I'm aware is a bit of a blindspot for me (when reading many moves before I'm thinking this space is an eye and this space is an eye and don't think about that some stones are in self-atari in the future).

[Uberdude]

So I played a 20s game on my 1060 with #188, so that's about almost 2k playouts per move, plus an extras from pondering and reuse from previous moves when things go as expected. It played an atari for a ladder which didn't work in another 3-3 invasion joseki, and when the 8d pulled out then it had enough playouts to realise the ladder didn't work and compromised for a bad result, but then a dozen moves later was winning again (wth was white e10?! says q10 would keep him at 60%). My understanding of this joseki is if the ladder doesn't work black's hane at n18 is questionable as it's white right to choose this variation and if black can't ladder it's hard to deal with the cutting stone (moreso as white already has r12). But reading the ladder back at the time to play n18 or another choice like p18 seems to be beyond LZ 188 with this many playouts. Another interesting thing: before playing the o8 ladder LZ was looking at forcing with r8 first (higher winrate but low playouts), and as it happened with the game continuation that would've been a very nice kikashi.

[sorin]

wth was white e10?!

Misclick?

[johnsmith]

@Uberdude
I noticed that in many of your games there is at least one ladder issue. Have you considered trying modification of LZ that knows ladders? It recognizes ladders in less than 50 playouts and uses a bit of cpu for that recognition (and gpu is used for position evaluation). The network itself is 20 blocks and 192 filters. It has ~60% winning score against the strongest 15 block network. All credit goes to ttl (github), jio (discord), ancalagon (discord). It works with Lizzie v0.5 (not v0.6 because it was built before the latest release).
You can download it here: https://we.tl/t-VKZh5CQq9d (wetransfer.com)

Example of your latest game:

sshot-1.jpg (169.78 KiB) Viewed 19528 times

[Mike Novack]

At this point, experimentation with the neural nets seems centered on "zero" versions of training. We need to keep in mind that the goals/objectives of neural net experimenters is not necessarily to as soon as possible come up with the strongest go playing nets as it is to study how nets learn something as difficult as go.

It is am interesting property of neural nets that prior (even damaged prior) training affects subsequent learning/(re)learning.

The open question is what WOULD constitute "ladder training". And at what stage of zero training should the zero training be interrupted for a course of "ladder training"