Leela Zero analysis of 'Making good shape' and other books.

[Uberdude]

For problem 7, it's hard to find a realistic position as that looks like the far approach to 5-3 joseki, but then black made an unusual block on 2nd line instead of tenuki or extend or jump, and then white tenukid (could be reasonable to a big place if black's block was stupid and slow) rather than defend corner shape or pincer. Here's a recent pro game with a pincer, LZ says block was slow (prefers extend), prefers defend corner with kosumi to black's pincer, and agrees with the attach rather than cut across knight move (would capture not connect when white ataris). The pincer obviously makes quite a difference here, for example with the book sequence the peep at p6 is annoying.

Or this Dosaku game if you make Dosaku play lower left shimari instead of kosumi defence (which LZ agrees with) then LZ wants to play the same attach instead of sacrifice. The peep (o5 here) after the gote wall seems to be a key minus of the book sequence.

Edit: about problem 107, I wonder what LZ thinks of p6 same shape at move 50 in this game from daal: viewtopic.php?p=227310#p227310

[John Fairbairn]

If we start with an empty board and say: "Black to play," LeelaZero offers just two moves (4-4 and 4-3). But many top pros will happily play 3-3. This month alone (August 2018) at least 15 pros have played this. And other AI bots have played it recently, too (e.g. Golaxy beat Ke Jie with it).

We believe there is a margin of error for all bot predictions (early in the game, at least). How substantial is matter of debate, but the fact that it is mentioned so often suggests many people regard it as too important to overlook.

In the case of the problems here, we get distortions because of the lack of stones elsewhere (not to mention that the question in the original is not necessarily "what is the best move" but "how does Black...."). We appear to have no sense of scale of this distortion.

Even with a full-board problem we can get distortions. Komi may have been different from 7.5. The answer may have been provided by a ghost-writing amateur, etc.

Then we have possible distortions due to style. We don't even properly understand the style of bots yet.

And of course bots are human - they are written by humans who make programming errors. They fall off ladders, etc.

I accept that over a large number of games, bots will beat humans almost every time. But is that because every move is better than a human's or just because fallible humans make the last big mistake through a lapse of concentration?

All of those points will have already been raised or considered by others. But for me what they amount to is just a huge question mark over any bot judgement. I have absolutely no feel for which factor counts more than others or by how much. In short, I know there is a question mark but have no idea what it means. I am like the caveman terrified by a lightning bolt.

There are several people here, however, who work with numbers or statistics and who presumably do have some feel for what's going on, what we can trust, what we can laugh off. What do they make of it all?

[Knotwilg]

To answer John, yesterday I did a separate exerice as Zermelo with some other 6d teaching material on whole boards (may publish later) and decided to continue, playing the "best move" each time. What you then see is still a wild fluctuation of the win rate, as LZ discovers new lines down the road which it had previously pruned. At one point the winning probability flipped sides while it had been at 70% on another.

Conventional wisdom averages out. AI averages out but differently. The former seems to care more about local efficiency, the latter about global sente. Clearly AI values sente a lot but in some cases only discovers down the road what sente delivers, which is no surprise. Conventional play is more honte-ish because it is less optimistic/certain of what it can do with the initiative.

I don't think we're at a point yet where AI can convincingly decide on textbook moves. But surely they lead to unbiased and very interesting arguments. Like Bill says, we want to reread a problem as "IF you want to connect ... but here's a way of playing that focuses on sente instead of connecting"

[Bill Spight]

To answer John, yesterday I did a separate exerice as Zermelo with some other 6d teaching material on whole boards (may publish later) and decided to continue, playing the "best move" each time. What you then see is still a wild fluctuation of the win rate, as LZ discovers new lines down the road which it had previously pruned. At one point the winning probability flipped sides while it had been at 70% on another.

Wow! What setting did you use? The more visits, the less chance of overpruning, I suppose. Or for analysis, there's the LZ version modified to do less pruning. https://github.com/AncalagonX/leela-zer ... v0.16-next

[zermelo]

I've thought about what useful this exercise actually teaches, and it maybe comes down to this: If a stronger player tells that some position is better than another, but you cannot with reasonable effort understand why, then just ignore that opinion for the time being. There's a good chance that he/she/it is wrong, and in addition, if you don't understand the reason, you probably cannot learn anything generalisable about the case anyway.

I don't know if this is something controversial or something that most people agree on already. At least for me it is a new way of thinking. I thought that a way to improve my positional evaluation is to use the judgement of stronger players as 'training data'. Now I feel I have used a significant portion of my study time just to learn some noise instead of the signal.

But of course you need to learn from _something_. I'm quite sure that we humans cannot learn zero-style (though for a devil's advocate, has this actually been proven...?). For instance from this 'Making good shape' book we can train our 'policy network' by taking the book solutions as training examples (just don't take the 'failures' as negative examples). We can also learn how to succeed in some local goals like cutting, connecting, or capturing some stones (but only when we actually understand the specific tactics clearly).

Obviously this does not come only from this little study. I think similar conclusion can be made in general from bot evaluation of joseki and pro opinions.

[Uberdude]

I'm quite sure that we humans cannot learn zero-style (though for a devil's advocate, has this actually been proven...?).

I think we could, and with fewer games than computers as we are better at learning/generalising from small data sets. But LZ has taken 9 million games to get to where she is today and humans tend to die before they can play that many games. I've sometimes wondered what would happen if 2 people were stranded on a desert island with a go board (and just knew the rules) for years and just played against each other, what would their go look like?

Here's an analysis of problem 87 I just opened the book at. The stated aim is "How can black make thickness for his stones?". I thought this shape could result from a strangely ignored outside approach to 3-4, but pattern searching shows it as a 6-4 joseki. There is one pro game that fits with the hoshi extension: Onoda Takuya 3p (black) vs Murakawa Daisuke 8p in 2017. Onoda played the squeeze starting with q18 which I correctly expected to be the book answer as it's a classic example of good move order to get maximum forcing moves. LZ doesn't like leaving high corner moves open so slightly preferred tenuki to add a move there.

MGS p87.PNG (702.64 KiB) Viewed 16287 times

If I turn the lower right corner into a 4-4 then it wants to continue locally, but with the book wrong answer of just p17 atari and then r14 sente block (omit s15 sente, interesting) and then p12 knight connection (initially preferred o12) instead of p14 solid in book. The book sequence would leave black at 32.8% on this board, whilst LZ's is 41.5%. 1% of that improvement comes from not playing s15, but it's mostly the more efficient knight move defence of the cutting point.

MGS p87 a.PNG (208.92 KiB) Viewed 16287 times

How does LZ like the book sequence? It's actually pretty close behind. With the s15 atari and solid p14 connect black is 37.3% which is > 32 so LZ agrees the book's right answer is better than the book's wrong answer. However, LZ would prefer to not s15 and then make the same efficient knight move defence for a 40.3%, just a little worse than p17 atari version. It's easy to see that although the q18 sacrifice allows black to seal the top side on the 2nd line, it gives white a stone at p17, reducing the liberties of the p16 stones and makes a cutting point at o16. Despite that weakness LZ still doesn't want to solid connect, but it's logical that if you are going to solid connect, it's more okay to play the q18 sacrifice because solid connect reduces o16 cut bad aji.

MGS p87 b.PNG (220.37 KiB) Viewed 16287 times

So my conclusion:
- the increased efficiency of defending the cut with knight move instead of solid connect is most important issue and book ignored it (though it did say "make thickness" so maybe that can be interpreted to mean don't play the best efficient move but play more solidly even if it's worse)/
- closing off the top side isn't as important as black getting a strong wall at p17 without o16 cut (in this position at least, and giving black f17 doesn't change that).
- think before s15 atari (it's a classic tseuji to make white s14 gote, but if white's not going to s14 soon do you need to play it now? it loses a ko threat and some yose compared to t15 if black gets to play first there)

P.S. Without extra side stone pro games are about half and half between the sacrifice squeeze and simple atari, e.g. http://ps.waltheri.net/database/game/70665/

[zermelo]

I'm quite sure that we humans cannot learn zero-style (though for a devil's advocate, has this actually been proven...?).

I think we could, and with fewer games than computers as we are better at learning/generalising from small data sets. But LZ has taken 9 million games to get to where she is today and humans tend to die before they can play that many games. I've sometimes wondered what would happen if 2 people were stranded on a desert island with a go board (and just knew the rules) for years and just played against each other, what would their go look like?

Right. I suppose I just meant that zero type learning would probably be much less effective than learning from better players in some way. But OTOH, maybe practising tsumego or endgame in zero style could be more effective (for improvement) than spending time to learn to imitate pro fuseki.

[John Fairbairn]

We seem to have agreed that bots tenuki more than human pros, and seem to like sente (hence also forcing moves) more.

What else? It's only an impression, because running a bot to check a human problem is so tedious, but it seems to me that bots generally agree that the humans have good suji - they get the flow of stones right - but they often get the first move wrong. In one problem I just looked at, LZ accepted an 18-move non-obvious (to me) pro sequence once move 1 was played - but it didn't like Black's move 1. The human comment was that was necessary otherwise Black ended up being counterattacked, but the bot was unfazed by the counterattack. From other examples, too, I get the sense that LZ evaluates the safety of groups quite differently from humans (not in a different way, necessarily, but certainly with different confidence levels). I'm guessing that it sees resources in reserve that humans either miss or are unsure about. In other words, it is effectively reading deeper.

Does that sound right to you? If it is, on the one hand it seems a little depressing - we'll never match depth even with a string of new proverbs - but on the other hand it means we can probably keep taking the suji pills prescribed by the pro teachers. It's mainly their evaluations we have to wary of.

Musing further, I suggest the only sane way to have some sort of confidence in your own evaluations (in the absence of obvious blunders, of course) is not to say "what would a pro play" but simply to ask "how consistent is this with my previous play?" That's essentially means choosing a style. We all think we do that, but in practice how often do we amateurs diverge, veering off to chase the clever bird that shoots out of the undergrowth, taking us away from its nest and we end up falling in a bog?

[Bill Spight]

I'm quite sure that we humans cannot learn zero-style (though for a devil's advocate, has this actually been proven...?).

I think we could, and with fewer games than computers as we are better at learning/generalising from small data sets. But LZ has taken 9 million games to get to where she is today and humans tend to die before they can play that many games. I've sometimes wondered what would happen if 2 people were stranded on a desert island with a go board (and just knew the rules) for years and just played against each other, what would their go look like?

Right. I suppose I just meant that zero type learning would probably be much less effective than learning from better players in some way.

I am of the school that leans towards learning from better players by playing against them. Not that you can't learn from a book or lecture, but most learning is caught, not taught.

But OTOH, maybe practising tsumego or endgame in zero style could be more effective (for improvement) than spending time to learn to imitate pro fuseki.

I have studied the endgame zero style and imitated pro fuseki. My study of tsumego has been a combination of both. IMX, imitation wins hands down for early learning. Up to where you have to start unlearning. In chess there are certainly geniuses like Capablanca, Morphy, and Fischer who did not rely much upon imitation but leaped ahead of everybody else.

[Bill Spight]

We seem to have agreed that bots tenuki more than human pros, and seem to like sente (hence also forcing moves) more.

I think you are addressing Uberdude, but if I may. . . . I agree that bots tenuki more, and so perhaps will play sente as part of that, or in preparation for tenuki.

What else? It's only an impression, because running a bot to check a human problem is so tedious, but it seems to me that bots generally agree that the humans have good suji - they get the flow of stones right - but they often get the first move wrong. In one problem I just looked at, LZ accepted an 18-move non-obvious (to me) pro sequence once move 1 was played - but it didn't like Black's move 1.

Interesting.

The human comment was that was necessary otherwise Black ended up being counterattacked, but the bot was unfazed by the counterattack.

Right. That's my impression of bot style, as well. What? Me worry?

From other examples, too, I get the sense that LZ evaluates the safety of groups quite differently from humans (not in a different way, necessarily, but certainly with different confidence levels).

Yes. This is why, I think, the bots don't often bother to make traditional human bases for their stones.

I'm guessing that it sees resources in reserve that humans either miss or are unsure about. In other words, it is effectively reading deeper.

Well, certainly the bots produce larger game trees than humans, both broader and deeper. (It is possible that Leela Zero reads too deeply, and not broadly enough. ) But it is not the calculation of variations that has made Leela Zero strong and keeps it improving, it is its evaluation of positions and choice of candidate plays. Improving them is what enables it to see better.

Does that sound right to you? If it is, on the one hand it seems a little depressing - we'll never match depth even with a string of new proverbs - but on the other hand it means we can probably keep taking the suji pills prescribed by the pro teachers. It's mainly their evaluations we have to wary of.

About the depth of reading of the bots. Their reading is global, and sometimes human local reading trumps it. A bot may judge a group to be safe while a human will see how to kill it or make a ko. This is also why, in the more nebulous cases where a bot will tenuki instead of making a local play to protect a group, I don't think it is that the bot has necessarily read the local situation more deeply than a human pro can. Usually I think it just has better judgement.

I agree that humans are not going to develop their calculation of variations much beyond what they are capable of at age 20 or so, but, just as Leela Zero and other bots continue to improve by improving their choice of candidate plays and their evaluation of positions, IOW, their judgement and intuition, humans can, too.

[Bill Spight]

Here's an analysis of problem 87 I just opened the book at. The stated aim is "How can black make thickness for his stones?". I thought this shape could result from a strangely ignored outside approach to 3-4, but pattern searching shows it as a 6-4 joseki. There is one pro game that fits with the hoshi extension: Onoda Takuya 3p (black) vs Murakawa Daisuke 8p in 2017. Onoda played the squeeze starting with q18 which I correctly expected to be the book answer as it's a classic example of good move order to get maximum forcing moves. LZ doesn't like leaving high corner moves open so slightly preferred tenuki to add a move there.

For one thing, I don't think that we have to strain after gnats in terms of Leela Zero's evaluation. All of these are possible. What about the keima to O-18? It closes off this entrance to the top side with sente, but without sacrificing a stone. OTOH, assuming that White plays Q-18, it gives White possibilities that the sacrifice does not, such as N-17.

If I turn the lower right corner into a 4-4 then it wants to continue locally, but with the book wrong answer of just p17 atari and then r14 sente block (omit s15 sente, interesting) and then p12 knight connection (initially preferred o12) instead of p14 solid in book. The book sequence would leave black at 32.8% on this board, whilst LZ's is 41.5%. 1% of that improvement comes from not playing s15, but it's mostly the more efficient knight move defence of the cutting point.

This keima connection is very nice. I would have chosen the book sequence, with the one stone sacrifice and the solid connection, but Leela Zero judges that to be bad. IOW, the value of this keima is worth more than sealing off the open skirt. How does the extension affect that decision?

How does LZ like the book sequence? It's actually pretty close behind. With the s15 atari and solid p14 connect black is 37.3% which is > 32 so LZ agrees the book's right answer is better than the book's wrong answer. However, LZ would prefer to not s15 and then make the same efficient knight move defence for a 40.3%, just a little worse than p17 atari version. It's easy to see that although the q18 sacrifice allows black to seal the top side on the 2nd line, it gives white a stone at p17, reducing the liberties of the p16 stones and makes a cutting point at o16. Despite that weakness LZ still doesn't want to solid connect, but it's logical that if you are going to solid connect, it's more okay to play the q18 sacrifice because solid connect reduces o16 cut bad aji.

My inclination would have been to make the solid connection because of the sacrifice, not the other way around.

So my conclusion:
- the increased efficiency of defending the cut with knight move instead of solid connect is most important issue and book ignored it (though it did say "make thickness" so maybe that can be interpreted to mean don't play the best efficient move but play more solidly even if it's worse)/

When I was 3 kyu or so I read a book on shape and it really opened my eyes. I became a shape player until I reached 4 dan. (I had to give up my attachment to good shape in order to do so. ) My personal definition of shape is local efficiency. OC, the keima connection is quite efficient, so I'm surprised they were tacit about it.

[John Fairbairn]

LZ doesn't like leaving high corner moves open so slightly preferred tenuki to add a move there.

Without intending to contradict that, I'd like to toss in question that could lead to a different interpretation.

In all the copious literature on thickness, I've seen next to nothing that talks about the timing of thickness. Thickness early on lets you make grand plans and maximise extensions (emptier board) but you are showing your hand very clearly and your opponent can dance around it. Also, thickness needs connections but early in the game you can't often tell whether a keima, tiger's mouth or solid connection is best, and this kind of thickness runs the risk of being overconcentrated. The converse more or less applies later in the game, but there is perhaps also greater risk of disruption.

So dies it matter - early or late? I ask this because among the many vague impressions I'm forming one is that bots don't like making thickness early on. Indeed they like giving it, as in the now notorious 3-3 invasion!

The Murakawa-Onoda game cited above was a case of very early thickness, and my impression was that Black just couldn't make that thickness do anything useful (he resigned early, too). In that light, the bot's choice of tenuki-ing to the shimari makes some sort of sense even to me. Do you think the bot's choice was not a hierarchy of one plan - make the shimari - but a hierarchy of two - 1. avoid early thickness, 2. make the shimari?

[Uberdude]

Going back to zermelo's points about how much we can trust our human experts, best way to learn etc. With super-strong AI it's now easy for weaker players to discover mistakes in what they used to take on trust from stronger ones, and some of the aura of strength will fade. But I'd still be happy to play as 'badly' as Rob van Zeijst 7d or some Japanese 9p from yesteryear who LZ says makes big mistakes nearly every move in the opening. In some specific situations where I have newer knowledge informed by more recent pro or AI thinking I can play better than them, but overall I'm still far weaker because the most important thing for one's strength is not how good your best moves are but how bad are your worst (and reading power). And I still think the best way to pull up the level of your worst moves is tsumego, so if I just wanted to get stronger I'd do that, but I play more for interest and fun.

I'm also reminded of a correspondence game I played on OGS some years ago, in which we played an avalanche variation and then another complicated joseki in the adjacent corner identically to a pro game (I'd looked it up and presume my opponent had too). Trouble was I was playing the side of the pro who lost, so I couldn't keep imitating for ever! So I studied the pro game with the aim of finding his mistakes so I could improve on his result (which might sound presumptuous, but I had his pro opponent to help). As it happens I managed this (with some help from my opponent) and now we've got LZ I think I'll analyze my game compared to the pro's in my journal.

Edit for posts made whilst I was drafting this:

John: Yes, I agree there is a tendency of bots to not be so impressed by early thickness, if we can call it that (as with the 3-3 invasion it's a wall but it's not yet alive). And they do seem very good at finding the right depth of reduction later. But I think with the shimari it is mostly about not liking high initial corner moves as I've seen this kind of wanting to come back to close it early in many positions, and it didn't mind making the thickness if the other corner was a 4-4 instead of "urgent to correct the mistake" 6-4.

Bill: If black plays the knight move to 2nd line LZ white would s16 capture to avoid r14 sente (p17 push to create cutting point beforehand possible) but definitely not q18.

[Bill Spight]

I still think the best way to pull up the level of your worst moves is tsumego, so if I just wanted to get stronger I'd do that, but I play more for interest and fun.

At this point I think that for DDKs the best way to pull up the level of their worst moves is to study damezumari at the end of the game and the lurking semeai. Big swings happen there. Otherwise, I would use a strong bot and see which moves lose the most percentage points.

Bill: If black plays the knight move to 2nd line LZ white would s16 capture to avoid r14 sente (p17 push to create cutting point beforehand possible) but definitely not q18.

Good point. Thanks. Does that hold true in response to O-17 as well? (O-17 and O-18 were not on my radar. )

[zermelo]

Going back to zermelo's points about how much we can trust our human experts, best way to learn etc. With super-strong AI it's now easy for weaker players to discover mistakes in what they used to take on trust from stronger ones, and some of the aura of strength will fade. But I'd still be happy to play as 'badly' as Rob van Zeijst 7d or some Japanese 9p from yesteryear who LZ says makes big mistakes nearly every move in the opening.

Yes, I agree with this completely, and that is why I suggested that we can still use the better players sequences as positive examples to teach our personal 'policy networks', maybe even when we don't understand the reasoning behind the moves. The bots have not changed anything in this regard.

What I think the bots have shown is that even the strongest players have been too opinionated about 'bad' positions even when they in fact don't understand them very well. Like the book in question that is full of 'failures' that are in fact perfectly valid sequences. Then the weak players like me have spent too much time trying to learn to replicate these judgements. I think we can learn something useful from this.