www.Ootakamoku.com - Modern fuseki practice.

[HermanHiddema]

Move A has been played 135 times, move B 14 times, move C 1 time. Now should move C be considered viable or not. Maybe it was a brainfart or a missclick or just some crazy experimentation that turned out to be really badly never to be tried again. B should almost definitely be still included as an answer, but why is it A played 10 times more often?

What if C was played in on of the 5 most recent games in the database?

What if B was played 10 times in the last year and 4 times before that, and A 20 times in the last year and 115 times before that?

Both such situations could indicate that a new move has been found, perhaps better, or just as good. Or they could simply indicate a change in style or fashion. Or, at those kinds of numbers, even the fact that a new pro has entered the scene and that he is very much a fan of a certain fuseki pattern.

I don't think there is an easy answer. I do think that finding any move played by a pro, even if turns out to be a pro-level brainfart, strongly indicates that at least your instincts are in the right place.

Anyway, thanks for the effort, it keeps improving and I think a lot of people will learn a lot from it.

[Ootakamoku]

This is partially what worries me. You're looking at a relatively small sample of games over a short time period. Very, very easily a move that is making up a low % of games might actually be the currently considered best move because once it was discovered and the results seen in a few games the previous move has been avoided until a counter to the new move is found. Going simply by frequency you could be punishing the only correct answer for this position currently considered and rewarding answers that are now considered inferior but were widely played in your sample beforehand.

There are 15000 games over 5 years. If a position has only a few examples, then it can hardly be considered common nowdays. And for a rare position the basis for the best move is actually just one or two pro players oppinion, those who were in the position within last 5 years. The move they played are clearly their oppinion of the best move for that position. Having another pro player vet the position probably wouldnt help since then we would just be exchanging one pro players oppinion for another. Yes there are a few positions which used to be considered standard but have now been "refuted" and he might help there. However for now my choice of using 5 years worth of games is an attempt to strike a balance, between having modern enough situations that there shouldn't be too many outdated fusekis, and the few that there are should be end up in the "maybe" category instead of "correct". While at the same time providing a diverse enough choice of possible moves for a position.

In actuality in most situations, there is only one correct move. Everything else leads to a loss. Well might be that regardless what move you play you will lose anyways, assuming both players play perfectly. Since we obviously can't go by those standards, as no one knows the right answer for most positions, we just have to settle with best available knowledge. If I were to ask a pro to vet positions we would just be exchanging the collective wisdom of pro players for one particular pro players oppinion. In truth no one can create a perfect set of situations with only the correct answers, it simply cannot be done. We are left with trying to minimize the errors and have positions and answers that are educational despite their imperfections. Just like you learn from studying pro games, even tho you know they make several mistakes during every game, the sum of the parts adds up to a positive value.

[Ootakamoku]

Move A has been played 135 times, move B 14 times, move C 1 time. Now should move C be considered viable or not. Maybe it was a brainfart or a missclick or just some crazy experimentation that turned out to be really badly never to be tried again. B should almost definitely be still included as an answer, but why is it A played 10 times more often?

What if C was played in on of the 5 most recent games in the database?

What if B was played 10 times in the last year and 4 times before that, and A 20 times in the last year and 115 times before that?

Both such situations could indicate that a new move has been found, perhaps better, or just as good. Or they could simply indicate a change in style or fashion. Or, at those kinds of numbers, even the fact that a new pro has entered the scene and that he is very much a fan of a certain fuseki pattern.

This is actually on the todo list. I intend to weight the pros answers according to how recent they are. So recent examples are valued much higher than older examples, despite having both in the database. While again not perfect, it should reduce the number of absolete fuseki variations.

[Boidhre]

This is partially what worries me. You're looking at a relatively small sample of games over a short time period. Very, very easily a move that is making up a low % of games might actually be the currently considered best move because once it was discovered and the results seen in a few games the previous move has been avoided until a counter to the new move is found. Going simply by frequency you could be punishing the only correct answer for this position currently considered and rewarding answers that are now considered inferior but were widely played in your sample beforehand.

There are 15000 games over 5 years. If a position has only a few examples, then it can hardly be considered common nowdays. And for a rare position the basis for the best move is actually just one or two pro players oppinion, those who were in the position within last 5 years. The move they played are clearly their oppinion of the best move for that position. Having another pro player vet the position probably wouldnt help since then we would just be exchanging one pro players oppinion for another. Yes there are a few positions which used to be considered standard but have now been "refuted" and he might help there. However for now my choice of using 5 years worth of games is an attempt to strike a balance, between having modern enough situations that there shouldn't be too many outdated fusekis, and the few that there are should be end up in the "maybe" category instead of "correct". While at the same time providing a diverse enough choice of possible moves for a position.

In actuality in most situations, there is only one correct move. Everything else leads to a loss. Well might be that regardless what move you play you will lose anyways, assuming both players play perfectly. Since we obviously can't go by those standards, as no one knows the right answer for most positions, we just have to settle with best available knowledge. If I were to ask a pro to vet positions we would just be exchanging the collective wisdom of pro players for one particular pro players oppinion. In truth no one can create a perfect set of situations with only the correct answers, it simply cannot be done. We are left with trying to minimize the errors and have positions and answers that are educational despite their imperfections. Just like you learn from studying pro games, even tho you know they make several mistakes during every game, the sum of the parts adds up to a positive value.

I don't mean this harshly, I hoping this is constructive criticism. What brings me to look at this this way was a lecture by Guo Juan that I wanted about the Small Chinese, she took time to point out several moves that would crop up as very common in the databases that were now not being played because of a flaw being found through research and pros moving to a different response in this position. The caution was that just because you can find many pro examples doesn't mean that a move is actually a good move in the fuseki. Rating moves through frequency alone will often work but almost certainly throw up some instances where you're marking as correct moves that (at least) high level players should be avoiding because they are known to be suspect now. Simple frequency analysis is a great start but you'll have to move past that to get closer to where you seem to want to go. Again, not intended as harsh just hoping to be helpful. Opening a discussion on how get an algorithm to spot these "common but refuted" moves as opposed to simple fashion changes might be an idea? (looking at oversampling from a single player's games or a single "school" of players might be an idea?)

[RBerenguel]

I don't mean this harshly, I hoping this is constructive criticism. What brings me to look at this this way was a lecture by Guo Juan that I wanted about the Small Chinese, she took time to point out several moves that would crop up as very common in the databases that were now not being played because of a flaw being found through research and pros moving to a different response in this position. The caution was that just because you can find many pro examples doesn't mean that a move is actually a good move in the fuseki. Rating moves through frequency alone will often work but almost certainly throw up some instances where you're marking as correct moves that (at least) high level players should be avoiding because they are known to be suspect now. Simple frequency analysis is a great start but you'll have to move past that to get closer to where you seem to want to go. Again, not intended as harsh just hoping to be helpful. Opening a discussion on how get an algorithm to spot these "common but refuted" moves as opposed to simple fashion changes might be an idea? (looking at oversampling from a single player's games or a single "school" of players might be an idea?)

Detecting fashion changes from refutations is probably very, very hard. As a matter of fact, detecting sudden (but minor) changes in continuous or discrete signals (in an automated manner, so you get an alert) is relatively hard (again, for minor changes you'd be able to eyeball but still wonder if it's that big of a change: you need fairly above-average techniques in signal processing just to detect small spikes or trends,) in a go board it would be almost impossible. Imagine having move A a keima, move B a 1-point jump. A has been played a lot, whereas B has been played recently. A players win 50%, B won his game(s) (given a sample size of, say, 3 games vs 50). How could you tell with this data (which is mostly what you'd get with automated analysis until go engines improve 5-6 stones) that B was a good innovation and not just a coincidence? Or even how to distinguish if it was just a change of fashion (Koreans attack, let's all play the keima! Japanese defend, let's play the 1-jump!) or a complete refutation?

[Boidhre]

Detecting fashion changes from refutations is probably very, very hard. As a matter of fact, detecting sudden (but minor) changes in continuous or discrete signals (in an automated manner, so you get an alert) is relatively hard (again, for minor changes you'd be able to eyeball but still wonder if it's that big of a change: you need fairly above-average techniques in signal processing just to detect small spikes or trends,) in a go board it would be almost impossible. Imagine having move A a keima, move B a 1-point jump. A has been played a lot, whereas B has been played recently. A players win 50%, B won his game(s) (given a sample size of, say, 3 games vs 50). How could you tell with this data (which is mostly what you'd get with automated analysis until go engines improve 5-6 stones) that B was a good innovation and not just a coincidence? Or even how to distinguish if it was just a change of fashion (Koreans attack, let's all play the keima! Japanese defend, let's play the 1-jump!) or a complete refutation?

I'm not thinking of a wholly automated system but one that'd flag a certain move for review or research. A was extremely common, now it's very rare and a new move B is being played, mark this for a look before it goes live as a problem or similar. Again you'd want to be pretty bloody strong to tell what's going on with the change but it's the best I can think of right now.

[RBerenguel]

Detecting fashion changes from refutations is probably very, very hard. As a matter of fact, detecting sudden (but minor) changes in continuous or discrete signals (in an automated manner, so you get an alert) is relatively hard (again, for minor changes you'd be able to eyeball but still wonder if it's that big of a change: you need fairly above-average techniques in signal processing just to detect small spikes or trends,) in a go board it would be almost impossible. Imagine having move A a keima, move B a 1-point jump. A has been played a lot, whereas B has been played recently. A players win 50%, B won his game(s) (given a sample size of, say, 3 games vs 50). How could you tell with this data (which is mostly what you'd get with automated analysis until go engines improve 5-6 stones) that B was a good innovation and not just a coincidence? Or even how to distinguish if it was just a change of fashion (Koreans attack, let's all play the keima! Japanese defend, let's play the 1-jump!) or a complete refutation?

I'm not thinking of a wholly automated system but one that'd flag a certain move for review or research. A was extremely common, now it's very rare and a new move B is being played, mark this for a look before it goes live as a problem or similar. Again you'd want to be pretty bloody strong to tell what's going on with the change but it's the best I can think of right now.

Oh, I see. Indeed... to know what it is it's probably very high dan or pro level. Or fuseki/joseki junkie like Oota is

[Ootakamoku]

I don't mean this harshly, I hoping this is constructive criticism. What brings me to look at this this way was a lecture by Guo Juan that I wanted about the Small Chinese, she took time to point out several moves that would crop up as very common in the databases that were now not being played because of a flaw being found through research and pros moving to a different response in this position. The caution was that just because you can find many pro examples doesn't mean that a move is actually a good move in the fuseki. Rating moves through frequency alone will often work but almost certainly throw up some instances where you're marking as correct moves that (at least) high level players should be avoiding because they are known to be suspect now. Simple frequency analysis is a great start but you'll have to move past that to get closer to where you seem to want to go. Again, not intended as harsh just hoping to be helpful. Opening a discussion on how get an algorithm to spot these "common but refuted" moves as opposed to simple fashion changes might be an idea? (looking at oversampling from a single player's games or a single "school" of players might be an idea?)

Last summer, I attended a go lecture by one of europes strongest players. He asked audience for moves, and I suggested a move. He told me my move was greedy and it wasnt played, then showed a way that would counter my move. However a few months later the move I suggested started to regularly appear in top pro games. Me suggesting it was just a coincidence as there was a lot more behind it than I could hope to comprehend. Simply dumb luck I happen to suggest the move at the time. What makes this interesting however is that this is exactly the kind of vetting by strong player. Yet in this case it provided the opposite result than you were hoping for. He advised me against a move that became a common fuseki variation later on.

It all comes down to having good enough information. No need to strive for perfection, it's just a waste of resources. I dont care if I occasionally get told a move is bad, when in actuality its good or the opposite. What matters is that the total adds up to something positive. If I were to teach a 5kyu, I would probably teach him 10% nonsense. Advising wrong variations to choose, reading situations wrong in a way he wont notice and there by arriving at wrong conclusions. 10% is huge number of mistakes if you think about it. Despite all that the 5kyu would gain alot from my 2dan advise. The pro game db, regardless whatever errors it may contain is still far more accurate than I could ever hope to be. If I were to offer a 5kyu a game review he would gladly take it, thinking its a good opportunity for him to learn from a stronger player. Yet when he is using the pro game db he is worried about errors that are almost nonexistant compared to any errors in my teaching, I find that puzzling.

[karaklis]

What of a situation which have occurred 150 times in pro games. Move A has been played 135 times, move B 14 times, move C 1 time. Now should move C be considered viable or not.

I think for playing strength up to 5d amateur it would suffice to use all moves that sum up to 95% of the total number of moves (be it one move or many moves, and irrespective of the time of the relative moves having been played).

Since if we have only 1 pro example from a position, but many answers from users its only reasonable to actually diversify the possible answers even if it comes at a slight expense in quality

I think in that case it is probably not necessary to use such a position as problem. The focus should be on positions that occur most often. This is where you (as a player) would benefit most from.

[Ootakamoku]

What of a situation which have occurred 150 times in pro games. Move A has been played 135 times, move B 14 times, move C 1 time. Now should move C be considered viable or not.

I think for playing strength up to 5d amateur it would suffice to use all moves that sum up to 95% of the total number of moves (be it one move or many moves, and irrespective of the time of the relative moves having been played).

Good idea, I mean the way to decide what to consider as viable out of the available information. Currently I have that if a move is played atleast 1:5 as often as the most common move its considered still correct.

[quantumf]

If I were to offer a 5kyu a game review he would gladly take it, thinking its a good opportunity for him to learn from a stronger player. Yet when he is using the pro game db he is worried about errors that are almost nonexistant compared to any errors in my teaching, I find that puzzling.

I agree entirely with this point, and in fact, the rest of your post. What I find, odd, therefore, is why you're focusing on fashionable moves? Perhaps you explained it earlier but I didn't understand it properly, but I'm not keen on being a 3d at fashionable fuseki moves, I'd rather be a 3d at good fuseki moves. So I'm unclear why just using the last 15,000 games is better than using all 70,000 games.

[hyperpape]

I'm not sure what concrete change this suggests, but I think it's important nonetheless: however you decide right and wrong answers for rank need not be the same as how you teach. Take the earlier case of a move that has 150 instances, 135 for A, 14 for B, 1 for C. I answer B. Maybe you say it's right, maybe you say it's wrong--I'm not sure.

But either way, you can somehow tell me that it's much less common. If I play C, perhaps you say I am wrong, but you can still tell me that a pro tried it. That tells me something, regardless of how I take it.

I'm not literally saying that you need to give that information (it might be clutter, it might be hard to automatically summarize, who knows). But I do think that comments on what the system should do should at least in principle distinguish between "this is how the rating system should see it" and "this is what the user needs to know."

[skydyr]

If I were to offer a 5kyu a game review he would gladly take it, thinking its a good opportunity for him to learn from a stronger player. Yet when he is using the pro game db he is worried about errors that are almost nonexistant compared to any errors in my teaching, I find that puzzling.

I agree entirely with this point, and in fact, the rest of your post. What I find, odd, therefore, is why you're focusing on fashionable moves? Perhaps you explained it earlier but I didn't understand it properly, but I'm not keen on being a 3d at fashionable fuseki moves, I'd rather be a 3d at good fuseki moves. So I'm unclear why just using the last 15,000 games is better than using all 70,000 games.

Well, you also don't want to get games from too early, or you'll see moves marked as 'correct' because, for example, the high approach to 3-4 was almost never played at the time. The trick is finding a good balancing point for this. Bill Spight (I think?) suggested going back 50 years earlier in this discussion, which seems like a reasonable starting point to get to the right time to me. Perhaps as a player gets stronger, the amount of time you go back should be slimmed down, perhaps to 15 years, to throw out a number. Similarly, maybe that shin-fuseki era move is fine for a 7 kyu player, but will show its weaknesses to a stronger one.

[Ootakamoku]

Well, you also don't want to get games from too early, or you'll see moves marked as 'correct' because, for example, the high approach to 3-4 was almost never played at the time. The trick is finding a good balancing point for this. Bill Spight (I think?) suggested going back 50 years earlier in this discussion, which seems like a reasonable starting point to get to the right time to me. Perhaps as a player gets stronger, the amount of time you go back should be slimmed down, perhaps to 15 years, to throw out a number. Similarly, maybe that shin-fuseki era move is fine for a 7 kyu player, but will show its weaknesses to a stronger one.

I had this discussion a few weeks back with some high dans. For them it seems the reasonable balance to be somewhere between 5 years and 10 years, depending who you ask. Reason I went with 5 years for now is that as the service is still in development so I don't want to grow the database too large too fast, until I'm certain of what I need.

[Ootakamoku]

I think I might need to add some easier tsumegos too, the current one are kinda urm.. difficult. Just need to find a source of lots of good low level tsumegos.