Is this opening really so bad?

[kvasir]

A lot of people seem to have an instinctive negative reaction at quantifying things with a number[...]

And I also think some people go too far in the other direction, considering such numbers to be almost meaningless. I think they are meaningful! [...]

This discussion reminds me of how some think komi is meaningless for most players and others think it is a huge deal for everyone. It is impossible to reconcile because it is really only an argument about degrees of hyperbole. The effect of komi can be measured but it doesn't resolve anything because the hyperbole can still be argued about.

Besides qualifying the score estimate as totally, almost, mostly, somewhat, usually, or never meaningless (or meaningful), one could ask instead "how good are these estimates?". The answer is probably different depending on if it is how good are they for the program that made the estimate, for other programs, or how good are they for a human player.

[John Fairbairn]

In any event, my impression (and I think pros may view this in a similar way) is that the bot is not measuring territory at all, but rather is measuring the initiative.

If I may be allowed to quote myself, what lightvector says (and thank you for responding) seems more or less to be the same as the above when he says:

I like the phrasing "the bot prefers white by 5-6 points" a lot more than "white is 5-6 points ahead". The former again makes it clear that it's a preference by the bot, rather than an objectively accurate calculation of the value of a position.

If you have the initiative (and I don't mean sente) you have control of where the game is headed. If you prefer a position, your preference must normally be, in some way, because you likewise have control of the game.

A lot of people seem to have an instinctive negative reaction at quantifying things with a number

POSITIVE instinctive negative reaction, perleaase...

I think for many people this is in part because in common parlance, giving a number like this connotes more precision and certainty than there is

Surely the problem is with the people who have a positive reaction to numbers and then over-interpret them or over-obsess about them. For starters, if you put a decimal point in an evaluation number you quote, you must be over-obsessing, yet I see decimal points here all the time. When active here, Bill Spight was a superior mathematician but knew when to squash nasty creepy-crawlies like decimals. I can recall a time when people used to say things like I'm 2.3 dan.

If you forget about AI, and instead took a friendly pro willing to patiently humor whatever questions you had for them, and you asked them which side they would prefer if we gave black an extra 2 points, 4 points, 7 points, etc... at some point they would probably switch from preferring white, to being uncertain, to preferring black.

This mirrors a famous story about Kobayashi Koichi, who was willing to switch sides, as I recall, if you gave just him one extra point.

[RobertJasiek]

Bill Spight has also explained and researched in infinitesemals. Small numbers and fractions are good if they are meaningful and correct. Then use them - otherwise, approximation is an option. A Prussian king was wrong when declaring pi to be 4.

[gennan]

Is this opening really so bad?

I have played this many times - both as black and as white - in low handicap games, but also in even games when san-ren-sei was still commonly played up to low-mid dan level.

This old joseki may be a bit slow for black, but if you prefer outside influence over territory, then I wouldn't advise against playing this with black.

Noguchi Motoki 7d EGF once gave a lecture in a go training camp. He showed a game of his where he played quite unorthodox, and still won convincingly against an opponent of his level.
His advice to the audience (SDK to high dan players) was to play the type of moves and the type of games that you enjoy. If he can play freely at his level, then surely we can.

[Knotwilg]

I've always explicitly understood the score as "the, or some kind of weighed, average of the outcome of the current number of games taken from this position by the neural network" not as "the score given perfect play by both". We know it can't be the latter because Go is not solved. When partaking in the lottery, I don't expect my gain to be exactly equal to the Expected Value of my gain either.

It's good to be reminded of that though because it does loosen the idea of the "best move" as being preferred by x points over other moves. I may have made such implication in the past, ignoring it was a shorthand for a difference in averages.

It's also good to understand that the "imprecision" of the blue move, being the difference between the value of the blue move and its successor blue move, also comes from different sets of evaluations.

And it explains why an initial advantage can grow or shrink over time, even if you play "the best move" every time. It's not that there are hidden pieces of intelligence accumulating over time, it's just the residu of randomness in the system.

(correct me if I'm wrong again)

[Kirby]

In any event, my impression (and I think pros may view this in a similar way) is that the bot is not measuring territory at all, but rather is measuring the initiative.

If I may be allowed to quote myself, what lightvector says (and thank you for responding) seems more or less to be the same as the above when he says:

I like the phrasing "the bot prefers white by 5-6 points" a lot more than "white is 5-6 points ahead". The former again makes it clear that it's a preference by the bot, rather than an objectively accurate calculation of the value of a position.

If you have the initiative (and I don't mean sente) you have control of where the game is headed. If you prefer a position, your preference must normally be, in some way, because you likewise have control of the game.

I think "initiative" is a good thing to have in go, but what does it mean, exactly?

Let's consider a 9-stone game between a 15-kyu player and a pro. It's, let's say, 20 moves into the game. The board position is still good for black, but the pro has been making strides toward closing the gap. Who has the initiative?

I would imagine that, at least in some cases, it's the pro who has the initiative here. With the wording used here, the pro has "control of where the game is headed" - if not from the board position, at least from the dynamic flow of the game.

However, if we are to analyze the board position at that particular point in time, the bot evaluation may very well be "B+7" or "B+8".

I'd personally consider white to have the "initiative" here, but for the bot evaluation to still favor black. They are different measurements.

That being said, I don't have a precise definition for "initiative", so some may not consider white to have the initiative in the hypothetical example that I am giving here. My current feeling for the word might be somewhat related to "momentum".

[mycophobia]

Noguchi Motoki 7d EGF once gave a lecture in a go training camp. He showed a game of his where he played quite unorthodox, and still won convincingly against an opponent of his level.
His advice to the audience (SDK to high dan players) was to play the type of moves and the type of games that you enjoy. If he can play freely at his level, than surely we can.

I've forgotten where I first read this but I treasure it as some of the best Go advice I've ever received

[kvasir]

And it explains why an initial advantage can grow or shrink over time, even if you play "the best move" every time. It's not that there are hidden pieces of intelligence accumulating over time, it's just the residu of randomness in the system.

I don't think you are "wrong" per-se but maybe the conclusion is too strong or the premise too simple.

The error of an estimate is not only due to the variance of the estimator but also the bias of the estimator. Bias of an estimator is the difference between the average estimate and the true value. You account correctly for the case when the estimator is unbiased but very likely this complex estimator that is Go AI is biased in ways that depends on the position (and other factors such as the number of playouts).

If for example we start in position of type A where the estimator is biased against moyos and then the game evolves into position of type B where the bias is close to zero, then it could be expected that this favors one player (the one playing with the moyo) in terms of how the estimates change and maybe in terms of actual game results. An example of type of position with bias that is close to zero would be the endgame! It might be optimistic to think this can be detected easily or that the effect is significant but I remember we discussed this in a thread about old videos on how to become a dan player

Another way that the estimate can drift is when the estimate in a position of type A has high variance but the game evolves into position of type B that has low variance. In this case the estimate in the first position will likely have a larger error than in the second position, if the stars align (which might not be that unlikely) then this can look like one player was favored and one could draw incorrect conclusions (about mistakes in the game or that the AI estimate favors something that it actually does not).


=Edit I wrote it favored the player playing against the moyo but on reflection it is the other way around and then that probably also favors that player in terms of game results. So I just changed it, hoping the point comes across.

[Gomoto]

gobank.jpg (33.86 KiB) Viewed 7122 times

[Gomoto]

PXL_20220915_225143376k.jpg (33.62 KiB) Viewed 7124 times

[Gomoto]

embrace the numbers
and
praise the emergent joy of go

[Elom0]

We have to understand what we're talking about . . .

In both chess and go, there is actually no such thing as winning probability. A game is either 100% winning for black, 100% winning for white, or a draw. This means means that any such concept of a winning probability is purely a concoction of the human brain or AI processors inability to know the perfect sequence of play. Winning probabilities are entirely a personal thing, to either the AI or the computer, and it doesn't make sense to nail yourself to another person or computers winning probabilities, regardless of strength, because each person has different learned styles. And even the style you've currently learned to play should be double checked, because much of if may be based on poor learning that left you playing with a style that is not in line with your natural, innate style based on your phenotype--or rather your genotype, since your phenotype may also be suffering from the same issues as your learned style. Tami is absolutely accurate that you must focus on style first before skill level. skill is style since style limits skill.

Of course, the above conclusion assumes that when we say 'winning' we're referring only to winning probability. In actuality, in Chess and Go, the control you have on the board when you win is half of what's meant. In chess it's square control, and potential square control represented by material, in go it''s territory or maybe area in general.

So if your training an AI to win by the maximum amount of territory it can, the purest way to do so is for it to multiply the probability of a position occurring by the score difference. So a position in which you win by 6 points but have only a a 50% chance of reaching is worth 3 points. If the is a semeai in which to a certain perspective the result is either a 50% chance between the game ends in a draw or one player wins by 50 points, then from that certain perspective that player is ahead by 25 points. But actually, what about the perspective of the player who's winning? Losing? The weaker player, but rather maybe the stronger players perspective is what we should take as who's winning. In essence, if we base who's winning purely on the likelihood to win the game, it's an absolute myriad since that means when an even game begins between a stronger player and a weaker player, by definition the stronger player is already winning and the weaker player losing, and that's not what we mean by a person being in a winning position. The alternative that works is if you assume perfect play and whether someone is winning or losing is in absolute terms, but then does away with the concept of winning or losing by a little or a lot, so that has to be applied on top of this concept, which kind of brings you back full circle. Note also AI winning probabilities aren't necessarily AI winning probabilities.

Let's you were watching the women's continental football tournament. One team has more points on the board, but at the point you're watch, the team with less goals changed their strategy, got motived is actually now playing way better to the point where it's obvious their going to win. So who's winning?

All this is says that if an AI or a pro says don't use so-and-so opening because for games at their level their is a slight disadvantage--you can be assured you needn't lose any sleep if that's what you use. In fact, I'd advise you to use this rule of thumb; anything less than first move advantage, in chess and go, is definitely on the table! So we should spend time stressing over AI's opinions and AI 'cancel? culture' where entire joseki's are wiped out because the AI deems it loses by 0.2 percentage points--okay I'm obviously exaggerating, but still, switch that time to thinking about what it means to win in go. The exact same applies to the social culture as well, ironically both became prevalent after March 2016 so it's interesting.

[Kirby]

In both chess and go, there is actually no such thing as winning probability. A game is either 100% winning for black, 100% winning for white, or a draw. This means means that any such concept of a winning probability is purely a concoction of the human brain or AI processors inability to know the perfect sequence of play.

I'd hypothesize that a similar argument could be made for most expressions of probability, as probability is a measure of uncertainty.

In the real world, if you roll a six-sided die, and have full knowledge of every physical dynamic of the world (gravity, hand friction, effect of the atmosphere, etc.), you would know with 100% certainty the result of the roll.

But that stuff is too complicated to quantify, so we have a simplified model, measuring probability in the absence of all of those minute real world variables.

[Elom0]

In both chess and go, there is actually no such thing as winning probability. A game is either 100% winning for black, 100% winning for white, or a draw. This means means that any such concept of a winning probability is purely a concoction of the human brain or AI processors inability to know the perfect sequence of play.

I'd hypothesize that a similar argument could be made for most expressions of probability, as probability is a measure of uncertainty.

In the real world, if you roll a six-sided die, and have full knowledge of every physical dynamic of the world (gravity, hand friction, effect of the atmosphere, etc.), you would know with 100% certainty the result of the roll.

But that stuff is too complicated to quantify, so we have a simplified model, measuring probability in the absence of all of those minute real world variables.

Yes, I should have phrased it instead as probability is an inherently subjective viewpoint of a physical system. Unless of course one believes in the fundamentality of quantum indeterminism, but determinism is the theory that most follows Occam's razor. From a perceptual sense however outside of mathematics all is probability.

[Elom0]

Baduk TV often has the opinion of 4 AI's running simultaneously, which is also how I think it should be used.