All Things Go Podcast

[00:01] Travis: This episode is sponsored by Baduk Pop, SmartGo One,

[00:04] Travis: AI Sensei and Go Magic.

[00:08] Travis: I’ve been on a bit of a rank plateau on the Online Go Server,

[00:13] Travis: and if you’ve listened to this podcast before, you hear me speak quite a bit about the fact that rank shouldn’t be as important as I think it ends up being for myself and for a lot of us.

[00:26] Travis: But I’ve gone through this period where I’m plateauing. I’ve just stuck right at around 9.5 kyu,

[00:33] Travis: and I’ve only been setting my correspondence games to playing people that are in the 8-10 kyu range.

[00:42] Travis: It’s just so bizarre. I think there’s so much variation in the consistency of play at this level, as I think has been documented in online forums by others that I’ll play players that are really strong or really weak, but they’ll still be at a given rank.

[01:00] Travis: And it’s just so interesting to not see any change there.

[01:06] Travis: Now, of course, the other side is that there are so many different aspects to playing this game that you could absolutely have two players that are at a same rank, which is sort of like an averaging of all those different sets of skills,

[01:18] Travis: but there’s still big swings because you have different strengths. So I totally get that, and I’m aware of that. But even still,

[01:26] Travis: I’m feeling like after 50, 60 games of playing over the last however many months,

[01:31] Travis: it’s just such an interesting thing that you keep hitting up against this sort of bottom rank,

[01:38] Travis: and it makes me wonder, is there any sort of behavior where all of a sudden you like, break out all of a sudden and bust through that plateau to a new level?

[01:49] Travis: I don’t know.

[01:51] Travis: Yeah, we’ll see. I don’t know that I’m terribly worried about the fact that I’m staying flat. Of course I want to improve my rank, and there’s been some players where along the way I beat them,

[02:01] Travis: and now I look at them and they’re like significantly higher in rank. So I’m like, oh, wow, that’s great. I,

[02:07] Travis: you know, kind of pipped one of these better players, but statistically it probably doesn’t mean anything.

[02:14] Travis: Just some thoughts rumbling around in my head there.

[02:19] Travis: This is All Things Go.

[02:36] Travis: I’ve seen this question come up over and over again in various online forums,

[02:41] Travis: and I’ve also had a lot of questions. And so I posted on this about wanting to speak to somebody that could speak intelligibly about Go rating systems and the differences between them and how they work.

[02:55] Travis: I’m really, really Thankful because a gentleman named Nikola reached out to me and said, hey, I’m a mathematician.

[03:03] Nikola Tsarigradski: This.

[03:04] Travis: This is my background, and I’m a Go player, but purely as a hobby,

[03:08] Travis: I like thinking about and looking at Go rating systems. So I think a lot of you will really enjoy this conversation. It was incredibly informative for me.

[03:20] Travis: And so we’re going to start that conversation.

[03:23] Travis: I hope you enjoy.

[03:28] Travis: Nikola, tell me a little bit about how you got into Go and how did you develop the special interest in Go ratings?

Interview with Nikola Tsarigradski: How Go Rating Systems Work

[03:36] Nikola Tsarigradski: Yeah,

[03:37] Nikola Tsarigradski: I think my interest in Go ratings are like, a little bit independent of me being interested in Go.

[03:42] Nikola Tsarigradski: I’m part of that generation that started after AlphaGo.

[03:46] Nikola Tsarigradski: I enrolled in the university in 2017,

[03:48] Nikola Tsarigradski: just a year after AlphaGo, and I started studying cognitive science. I was doing like, a lot of math, AI stuff. So this was just like,

[03:56] Nikola Tsarigradski: fresh, big news, like us having gotten a superhuman AI.

[04:00] Nikola Tsarigradski: And my university had a Go club, so I just joined it.

[04:03] Nikola Tsarigradski: And I don’t know, for the first few years, I was just hanging around at 20 kyu, having a pizza with people and chatting, but at some point, some of my friends started to actually improve, and then I got deeper into it.

[04:17] Travis: Okay, got it. And then your background is your mathematician?

[04:20] Nikola Tsarigradski: Yes, I’m about to graduate with my master’s.

[04:24] Nikola Tsarigradski: I guess this is also kind of why I’m interested in rating systems.

[04:28] Nikola Tsarigradski: To be fair, I.

[04:29] Nikola Tsarigradski: I do feel, like, a little bit insecure because I. I feel like some of your previous guests on some topics have been having, like, far more expertise in what they were talking about in me.

[04:38] Nikola Tsarigradski: Like, to me, this is just something like I’m. I’m interested in.

[04:41] Nikola Tsarigradski: I played chess before I started playing Go, and already there I. I thought that, like,

[04:46] Nikola Tsarigradski: I don’t know, like from a mathematical perspective or statistical perspective, rating models are just interesting. I don’t know. I always enjoy it,

[04:54] Nikola Tsarigradski: reading up on them. And with go, it gets a little bit more complicated because in Go,

[05:02] Nikola Tsarigradski: there’s a bunch of things that you desire from a rating system,

[05:06] Nikola Tsarigradski: but some of them preclude each other. So it’s actually pretty difficult to design a system that gives you everything you want.

[05:14] Travis: Interesting. Okay, so let’s start with the Elo system, because that’s what’s used in the chess world on the Go side of things. I know that they’ve. In certain servers, like the Online Go Server, you could actually show what the Elo calculation is.

[05:28] Travis: So how does the Elo system work in sort of plain language?

[05:32] Nikola Tsarigradski: I’ll try to explain it as simple as possible.

[05:35] Nikola Tsarigradski: You have basically A probability distribution of how likely is it that someone will beat someone else.

[05:44] Nikola Tsarigradski: And when you look at two players,

[05:48] Nikola Tsarigradski: you can express this probability with a number. And you can basically translate the winning probability between two players into a rating difference. This is like the basic belief of the Elo system.

[06:03] Nikola Tsarigradski: When you have two players that are 100 rating points apart,

[06:07] Nikola Tsarigradski: this is basically defined to have a 63% winning chance for the stronger player. Or I guess in chess it’s expected outcome of 0. 63 and it also draws.

[06:21] Nikola Tsarigradski: Okay, so basically if they play like a thousand games, you would expect stronger player to get 630 points out of thousand.

[06:29] Nikola Tsarigradski: Got it.

[06:30] Nikola Tsarigradski: In the next step,

[06:31] Nikola Tsarigradski: you can use these predictions to also calculate the new rating based on game result.

[06:42] Nikola Tsarigradski: So if you have two players that are 100 Elo points apart, so the stronger player wins 63% of the time,

[06:50] Nikola Tsarigradski: you would also want your rating to converge to this 100 point difference. If people kept playing. So assume that one of the players is misrated,

[07:01] Nikola Tsarigradski: then you could just keep playing games over and over again. And if 100 Elo difference is the true rating difference of the two players, the rating would eventually end up exactly there.

[07:13] Nikola Tsarigradski: I see.

[07:14] Nikola Tsarigradski: And this is basically like this fundamental like self correcting property that any rating system has.

[07:21] Nikola Tsarigradski: And with elo, it’s more or less the simplest.

[07:26] Travis: So within that Elo system,

[07:27] Travis: is there an aspect that incorporates how many games you’ve played. So Whether you’ve played five games or you’ve played 500 games,

[07:35] Travis: is there any part that affects that part of the rating?

[07:40] Nikola Tsarigradski: It depends. In the classical Elo system,

[07:44] Nikola Tsarigradski: I mean, it’s named after a Hungarian mathematician from I think the 60s or 70s came up with it.

[07:50] Nikola Tsarigradski: Within the original system, the answer is no. The Elo system is a direct incremental rating system.

[07:57] Nikola Tsarigradski: So you have only two numbers, which is your current rating,

[08:01] Nikola Tsarigradski: and then you have like a rating update which just updates your rating based on a result.

[08:06] Nikola Tsarigradski: So the classical Elo system does not take that into account. But it’s a very, very good point that you raised because if you think about how ratings develop,

[08:14] Nikola Tsarigradski: this is actually like a very important consideration.

[08:17] Nikola Tsarigradski: So it doesn’t matter. Like it’s something that you do want to take into account. When you, when you calculate your ratings,

[08:24] Travis: is there any sort of point of reference where it uses a sort of proverbial strongest player or strongest thing to say, this is the North Star, this is the thing or person that’s the strongest.

[08:37] Travis: Everything is in relationship to that or not really.

[08:41] Nikola Tsarigradski: This is an excellent question. Which I was hoping to get to.

[08:44] Nikola Tsarigradski: I was thinking maybe we’d get there later.

[08:47] Nikola Tsarigradski: I can just quickly answer it for elo. Elo. The answer is no.

[08:51] Nikola Tsarigradski: The Elo system is fundamentally untethered. Is that the right word?

[08:56] Travis: Yeah.

[08:57] Nikola Tsarigradski: Like, there’s no fixed level of skill that corresponds to a rating. All ratings are relative. Elo doesn’t make a difference between, like, grandmasters playing like 2800 to 2700 or like beginners playing 1000 to 900.

[09:12] Nikola Tsarigradski: It’s all the same.

[09:13] Nikola Tsarigradski: And it has this, like, massive weakness that the ratings are not comparable over time periods,

[09:21] Nikola Tsarigradski: because when the entire population shifts in scale, you cannot measure this with the rating system.

[09:28] Nikola Tsarigradski: Yes, but this is really, really relevant to go. And I’ll try to loop back to that later.

[09:34] Travis: That’s helpful. So we have the Elo system. This is used in chess, as you said. It’s a differencing in points,

[09:41] Travis: and it’s looking at a probability distribution.

[09:44] Travis: Then we have some of the other systems I’m aware of is the Glicko System or the Glicko-2 system.

[09:51] Travis: And I believe that’s used at least again on the Online Go Server system.

[09:56] Travis: How does that work? How is that different from the Elo system?

[10:00] Nikola Tsarigradski: The Glicko system is really, really similar to Elo system. It’s basically just a statistical improvement of the existing Elo model.

[10:10] Nikola Tsarigradski: So what you threw in earlier about does the rating system take, for example, into account, like, your most recent games?

[10:18] Nikola Tsarigradski: The Glicko system has,

[10:20] Nikola Tsarigradski: I hope I can list all of them,

[10:22] Nikola Tsarigradski: two or three improvements over the Elo system.

[10:25] Nikola Tsarigradski: The main thing it introduces is uncertainty.

[10:28] Nikola Tsarigradski: So you think that a player who just signed up for tournament, first time in their life, didn’t have a rating before,

[10:35] Nikola Tsarigradski: and their chess Coach says they’re 1400 here a day.

[10:39] Nikola Tsarigradski: Okay. Their chess rating is significantly more uncertain than the rating of someone who has a thousand games played.

[10:49] Nikola Tsarigradski: And the way you incorporate this into your rating system is that you have some sort of uncertainty measure. This is like a second number you introduce that depends on a bunch of parameters,

[11:00] Nikola Tsarigradski: and this affects your rating update value. So if your uncertainty level is very high,

[11:08] Nikola Tsarigradski: your game results are amplified. So if your uncertainty is really, really high, this means that if you win a bunch of games,

[11:15] Nikola Tsarigradski: it means that you’re probably not on a lucky streak, but actually underrated, so your rating increases quicker.

[11:22] Nikola Tsarigradski: And then if over a short amount of time, you play a lot of games and you’re performance within those games matches your actual rating,

[11:31] Nikola Tsarigradski: the uncertainty of your rating will go down.

[11:34] Nikola Tsarigradski: And so then if you have like bad day, you go to a tournament, you have five losses in a row, the rating system will recognize that probably you’re not just like really bad at the game right now.

[11:46] Nikola Tsarigradski: You just had a bad day. We’re pretty certain that your rating is not actually that bad. So it will affect your rating list.

[11:53] Nikola Tsarigradski: This is like the the main addition of the Glicko or the Elo system.

[11:58] Nikola Tsarigradski: And then the second one is that you also measure uncertainty as a function of time. As in like how many games do you play over time?

[12:07] Nikola Tsarigradski: For example, if you have your rating,

[12:10] Nikola Tsarigradski: the rating will not change if you don’t play games,

[12:12] Nikola Tsarigradski: but over time your uncertainty measure will increase if you don’t play.

[12:17] Travis: Got it? So if I play a game, the last game I played, I’m a certain Glicko rating and I don’t play for a year and then I play a game and I get a certain result.

[12:29] Travis: Let’s say I get a win,

[12:30] Travis: but then let’s say versus playing, you know, the next day,

[12:34] Travis: it’s going to have some time factor there that’s going to change the result.

[12:40] Travis: If we imagined like it was the same rank of person and I played them, you know, tomorrow versus a year from now,

[12:46] Travis: how would that affect it?

[12:48] Nikola Tsarigradski: If you play them today and you win as compared to if you play against them in a year after playing for a year and you win, you would get fewer rating points today.

[13:00] Nikola Tsarigradski: You will get more rating points next year because the model assumes that if you didn’t play for a year and you come back with the win, most likely you practiced in the meantime it got better.

[13:11] Travis: Interesting.

[13:17] Travis: This episode of All Things Go is brought to you by our sponsors,

[13:21] Travis: each of them helping you learn, study and enjoy Go in new ways.

[13:25] Travis: First up is Baduk Pop, the app that makes learning Go fun and accessible with interactive puzzles, bite sized lessons and an AI tutor to guide you step by step. It’s perfect for players on the move.

[13:37] Travis: You can download it or learn more at BadukPop.com that’s b a d u k p o p. com this episode is also sponsored by AI Sensei, your personal game review partner.

[13:51] Travis: Upload games and it instantly spots mistakes, offers stronger moves and helps you learn from each match.

[13:58] Travis: Give it a try at AISensei.com this episode is also sponsored by SmartGo One the All in One Go Toolkit for Apple devices.

[14:08] Travis: Replay professional games, solve problems and keep your SGF collection organized. All in one elegant app. I also really enjoy the library function and seeing dynamic diagrams of the actual game States in books.

[14:23] Travis: Check them out at SmartGo.com and then finally Go Magic, the interactive learning platform that makes Go come alive with structured video courses, exercises and lessons from top teachers. It’s an incredible way to accelerate your Go journey.

[14:39] Travis: I personally love their Ghost mode in the puzzle section of the website. It’s really great to try to practice moves without seeing them.

[14:48] Travis: You can find them at GoMagic.org that’s G-O M A G I C. org.

[14:58] Travis: And is there a sort of loading like a front loading in the Glicko or the Glicko-2 system? Like I know on the again OGS,

[15:07] Travis: I think they say that they more heavily weight.

[15:10] Travis: Is it the last 10 games or something like this that you’ve played in calculating the rating system or am I thinking about it incorrectly?

[15:21] Nikola Tsarigradski: Maybe I’ll make like a slightly bigger loop now. There’s many ways to classify it. There’s basically like four ways to do ratings.

[15:30] Nikola Tsarigradski: One way to do it is to have a simple rating,

[15:34] Nikola Tsarigradski: which we do not use in games like chess and go.

[15:37] Nikola Tsarigradski: This is something you do if you have like a very, very, very small sample size.

[15:42] Nikola Tsarigradski: In a chess world, it is kind of common to talk about performance ratings. Okay,

[15:48] Nikola Tsarigradski: so you just say like what was the performance rating in a tournament of a player which was basically the rating that would have won or lost zero points with that specific result.

[16:02] Nikola Tsarigradski: You play five players,

[16:04] Nikola Tsarigradski: produce a certain result, and then your performance rating is the rating that would not have changed your rating after those results.

[16:12] Nikola Tsarigradski: This would be like something like a simple rating, basically something that you only base on like a small sample size of data.

[16:18] Nikola Tsarigradski: This is something that’s not really relevant.

[16:21] Nikola Tsarigradski: The main three ways to figure out ratings in a predictive way are incremental ratings. This is everything that has to do with Elo or Glicko.

[16:34] Nikola Tsarigradski: It’s also most of the things are used in the Go world and the chess world and then you have a decaying-history rating.

[16:43] Nikola Tsarigradski: The idea of decayed history is that you do mainly take like the last few games you’ve had into account and you ignore the really, really old games.

[16:55] Nikola Tsarigradski: The advantage of this is that your rating will be more stable because I don’t know if you’re familiar with the website of the European Go database.

[17:06] Travis: Yes, I have seen it.

[17:07] Nikola Tsarigradski: You can look up player profiles and it shows you a graph how their rating has developed over time.

[17:14] Nikola Tsarigradski: And if you look up the pages of some really, really old players who have been like continuously playing over the years,

[17:20] Nikola Tsarigradski: you will notice that their rating is still like Going up and down like crazy.

[17:25] Nikola Tsarigradski: And overall, it stays kind of like within the same rank, but it still like moves up and down all the time.

[17:31] Nikola Tsarigradski: And that’s because if you have like a stable rating, like a stable skill, basically, that doesn’t change.

[17:36] Nikola Tsarigradski: Going to tournaments is still kind of like a random one.

[17:39] Nikola Tsarigradski: Like, you might play really, really well and win five games in a row to a degree. It’s like a random event.

[17:46] Nikola Tsarigradski: And the only reason it doesn’t diverge is that if your rating gets too high,

[17:51] Nikola Tsarigradski: you start getting penalized for losing against players of your own strengths. And if your rating gets too low, then you get disproportionately more points for winning against players of your own strength.

[18:03] Nikola Tsarigradski: Okay,

[18:04] Nikola Tsarigradski: but you still have this effect of your rating kind of looking like, like a stock price, just going like up and down like crazy.

[18:11] Nikola Tsarigradski: And we are modeling your skill at the game, and we assume that your skill in the game, your actual true skill, does not go up and down like this from day to day.

[18:20] Nikola Tsarigradski: It’s probably stable. And these effects of it going up and down like crazy are just artifacts of how we measure your rating.

[18:29] Nikola Tsarigradski: So the idea of the decaying-history rating is that you basically only take like a certain number of your most recent games into account and you weigh them less and less the more new games you play.

[18:44] Nikola Tsarigradski: So basically, like, you decay the value of the old games.

[18:47] Nikola Tsarigradski: And so your rating will be like an average of your most recent performance.

[18:52] Nikola Tsarigradski: So this means that winning or losing any one game at a time will not affect your rating much. Your rating will be more stable.

[19:01] Nikola Tsarigradski: And then there’s the method of taking this to the next level, which is doing a whole history rating.

[19:07] Nikola Tsarigradski: This is like the really, really fancy stuff where you basically take like the entire game history of a player into account. It also discounts old games, of course, like, the most recent games are most important,

[19:17] Nikola Tsarigradski: but it also takes into account the covariance of your performances.

[19:24] Nikola Tsarigradski: So, for example, if you lose five games in a row on one single day at one tournament,

[19:30] Nikola Tsarigradski: those are highly correlated. Like, most likely you had a bad day, maybe you lost a game and you fell into like a bad mood and you kept playing worse and worse,

[19:39] Nikola Tsarigradski: right? Whereas if you lose like one single game every single month probably means that you’re actually overrated. And so like the whole history rating approach,

[19:51] Nikola Tsarigradski: basically taking every single game you’ve played,

[19:53] Nikola Tsarigradski: the time sensitive context,

[19:56] Nikola Tsarigradski: also whether or not you keep playing against the same player or a different player,

[20:00] Nikola Tsarigradski: and using a very profound model with Bayesian inference to figure out what the most Likely rating is for you. So, like, which rating right now most accurately explains your previous results?

[20:17] Nikola Tsarigradski: And there is a website which is called goratings.org which lists ratings of the top 1000 Go players in the world,

[20:27] Nikola Tsarigradski: which I think, with the exception of two or three names, are all East Asian players. I think there’s like two or three European players there and maybe two or three American players.

[20:38] Nikola Tsarigradski: And that rating most famously uses whole history rating.

[20:43] Nikola Tsarigradski: I think it’s just like pet project of the person who’s entertaining the website.

[20:48] Nikola Tsarigradski: Right. And just using the data of game results to compute these ratings like they’re not official within the professional associations in East Asia.

[20:57] Travis: When I looked online, my understanding is that the Fox Go server might also use a whole history rating. I’m not sure.

[21:06] Nikola Tsarigradski: Yes.

[21:07] Nikola Tsarigradski: So stuff like Fox is, I think, very relevant because all of the rating system we’ve talked so far,

[21:13] Nikola Tsarigradski: they have worked on the assumption that we want our ratings to be predictive in some fashion.

[21:20] Nikola Tsarigradski: You don’t want your rating to just be like a number that you can compare to someone else and be like, hey, I’m better than my friends. But you actually want that number to carry some meaning for the game.

[21:29] Nikola Tsarigradski: And that is not, like, strictly necessary. You can have a rating system that doesn’t do that.

[21:35] Nikola Tsarigradski: One example of this is the Fox rating system. So the. The way the Fox rating system works is depending on which rank you are,

[21:44] Nikola Tsarigradski: you keep playing games. And as soon as you have a certain number of wins or losses within a certain window of games, you rank up or you ranked up.

[21:55] Nikola Tsarigradski: I think I’m currently 4 dan on Fox, and I have to have like a window of 20 games.

[22:00] Nikola Tsarigradski: And if within these 20 games at any given point there’s 14 wins,

[22:04] Nikola Tsarigradski: that ranks me up and resets all of my wins.

[22:07] Nikola Tsarigradski: And if there’s, I think, like 13 losses, I rank down instantly, and then the entire history is removed. And this is not a predictive rating system. It’s a very, very simple rating system,

[22:18] Nikola Tsarigradski: but it means that you, like, oscillate between ranks.

[22:21] Nikola Tsarigradski: You’re like, too weak for 4 dan. But too strong for 3 dan. And it doesn’t really, like, give you any intel on how you will perform against someone else who’s 4 dan, for example.

[22:32] Nikola Tsarigradski: It reminds me a lot of something, for example, that we use here at my Go club,

[22:37] Nikola Tsarigradski: comparing it to another Go club also in Finland.

[22:42] Nikola Tsarigradski: So my Go club here in the city of City in Finland, we have an internal rating system which I appreciate a lot because it is incredibly simple,

[22:52] Nikola Tsarigradski: but it makes a lot of sense because it provides us with what we need.

[22:56] Nikola Tsarigradski: Any player has like an integer number of points. I think I currently have 56.

[23:02] Travis: Okay.

[23:02] Nikola Tsarigradski: And this does not correlate to a rank. But whenever we have a game at club nights and we decide to have it like rated within the club,

[23:11] Nikola Tsarigradski: you get one point for a win, then you lose, one point for a loss,

[23:16] Nikola Tsarigradski: and then for each five points of difference there’s one handicap stool.

[23:22] Nikola Tsarigradski: So if I play the same player again and again and I keep losing, I’ll keep getting handicapped until like our games are evened out.

[23:30] Nikola Tsarigradski: Huh.

[23:31] Travis: Interesting.

[23:32] Nikola Tsarigradski: This isn’t really predictive,

[23:34] Nikola Tsarigradski: but it is self correcting at least. So if like your ratings are off, they will keep fixing themselves.

[23:41] Nikola Tsarigradski: And I really, really like this because it’s incredibly simple. We have like a small HTML website where we can just enter the results and it calculates everything.

[23:50] Nikola Tsarigradski: And since those games are not rated within the egf,

[23:53] Nikola Tsarigradski: there’s like absolutely no necessity to have a predictive system.

[23:56] Nikola Tsarigradski: As opposed to the Go club in Tampere, which is like the nearby city here they there have a system where they import the EGF ratings of everyone,

[24:07] Nikola Tsarigradski: I think once per year or twice per year and then they have like a fancy spreadsheet where every single time they play a game they calculate the official EGF rating update.

[24:18] Nikola Tsarigradski: They wait a little bit and then this changes like their top layer ratings. So you basically like you import your official EGF rating once and then when you play games at the club your internal club rating starts to like shift depending on the rules of the EGF rating system,

[24:34] Nikola Tsarigradski: but not officially reporting results to the egf.

[24:37] Nikola Tsarigradski: So this is like a very elaborate way of doing the club internal ratings. Whereas like we have like an extremely simple way of doing the club ratings. But if you have like a population of 20 players who’s coming to the club nets regularly,

[24:47] Nikola Tsarigradski: you don’t need a fancy predictive rating system at all. So like the rating system you use should match your needs.

[24:53] Travis: Hmm,

[24:54] Travis: that’s a really great point and I think we’re going to hopefully get into that in a minute on matching your needs and what the real goal is with a given rating system.

[25:03] Travis: I want to go back to something you said Nikola before about a rating being predictive.

[25:09] Travis: I want to better understand that because one question I’ve had when I look at for example the Online Go Server and their use of, I think it’s the Glicko-2 system.

[25:19] Travis: So for example, let’s say that I’m 8. 5 kyu plus or minus, right? Let’s say it’s 1. 0.

[25:28] Travis: And let’s say I play an opponent that is in the six range, right? So there are 6 kyu plus or minus.

[25:36] Travis: But let’s say it’s a larger deviation, a larger plus or minus.

[25:40] Travis: What I notice is that if I and playing that person,

[25:44] Travis: I look in their history and I see, oh, wow, they have not played many games on OGS. They’re a new player under this name at least,

[25:53] Travis: and I see they’ve only played 20 games or 25 games.

[25:57] Travis: Of course, that’s not a very large sample size. It’s a small one. And they’re showing in the 6 range of ranks or rating.

[26:06] Travis: But when I look at what games they’ve played and I look at what games they’ve won,

[26:12] Travis: they’ve not beaten a single player that is at the level of 6 kyu or better.

[26:19] Travis: It’s always been something lower. 11 kyu, 12 kyu, 14 kyu, whatever it might be.

[26:26] Travis: Is this the component of the Glicko 2 system that is predictive, where it’s saying, hey,

[26:33] Travis: you can beat somebody that’s at a rank that we’re not even going to give you, but because you’re winning in so many successive tries,

[26:40] Travis: we’re going to actually make your rating a higher one than who you’re playing. Am I making any sense?

[26:47] Nikola Tsarigradski: Yes. Your rating basically just predicts a win rate. And if you keep on playing bears who are exclusively weaker than you,

[26:57] Nikola Tsarigradski: your rating is still going to predict a win rate.

[27:01] Nikola Tsarigradski: So for like a 200 rating difference in Glicko, I think that’s right, 75%,

[27:06] Nikola Tsarigradski: it goes up. It’s basically dictated by a bell curve.

[27:12] Nikola Tsarigradski: So when the ratings start getting further apart, they get closer and closer to 100%.

[27:18] Nikola Tsarigradski: And this is why only playing much weaker or much stronger opponents is kind of a problem for the Glicko system,

[27:28] Nikola Tsarigradski: because when the win rates are extremely high or extremely low,

[27:33] Nikola Tsarigradski: the variance increases a lot. So if you play against player against, you have like a 99%,

[27:39] Nikola Tsarigradski: let’s say they’re like the true 10 kyu and you’re a true 2 kyu, let’s say, and then you have like a 99% win rate.

[27:47] Nikola Tsarigradski: The difference between you being a 5 kyu having like a 95% win rate or being a 3 dan and having a 99. 99% win rate is a difference that is very, very difficult to measure with a small sample size.

[28:02] Nikola Tsarigradski: You would need thousands and thousands of game to really measure this difference. If you keep playing against someone who’s significantly weaker or significantly stronger than you.

[28:11] Nikola Tsarigradski: So if someone is only playing against weaker players, then the predictive model will say, okay, they’re most likely like stronger than that, even though they’ve never beaten any one piece stronger.

[28:20] Nikola Tsarigradski: But how much stronger they are will be difficult to predict. So the rating might be pretty inaccurate. If you want to have like an accurate rating,

[28:29] Nikola Tsarigradski: the most efficient way to do it is to play against players who are similar strength.

[28:35] Travis: I see. So if you play against those theoretically weaker players.

[28:39] Travis: If I keep playing over and over against weaker,

[28:42] Travis: theoretical weaker ranks, let’s say that I’m truly a 3 kyu and I’m playing a whole bunch of 11 kyu players.

[28:50] Travis: If I keep winning over and over and over and I only play 11 kyu players,

[28:55] Travis: is anything going to happen from that perspective? Like if I played a thousand games and I win every single 11 kyu game, will the system have a really difficult time of landing at 3 kyu at some point?

[29:08] Nikola Tsarigradski: If you get for example to a point like for example, if you’re like a 2 dan and you play against an 11 kyu and you win,

[29:16] Nikola Tsarigradski: then the prediction of the result of the game is also corresponding to your rating update. So for example, if you have a 99% win rate and there are 10 rating points to be distributed,

[29:32] Nikola Tsarigradski: it doesn’t really work like this. But like let’s, let’s assume we have like 10 rating points that are being distributed. If like a 1 dan faces a 1 dan, then the winner can gets 5 points and the loser loses 5 points.

[29:42] Nikola Tsarigradski: But if 2 dan plays against an 11 kyu,

[29:45] Nikola Tsarigradski: then if the 2 dan wins, then

[29:49] Travis: will only get like 0. 1 rating. 0001 points. Yeah.

[29:54] Nikola Tsarigradski: Yes, yes. And then, and then the, and then the 11 kyu will like lose a tiny, tiny amount of. And the, the idea is that if the win rate is really 99%, you will do this 99 times and you will keep getting like this tiny, tiny, tiny rating and in one time you will have like the disastrous loss and you will lose all of the rating in like one go.

[30:15] Nikola Tsarigradski: Got it?

[30:15] Travis: It’s all downside. Not a lot of upside for the stronger player if they’re truly.

[30:20] Nikola Tsarigradski: I don’t know if you are familiar with this, but something very related to this has been causing a lot of drama in the chess world lately.

[30:32] Nikola Tsarigradski: Because in,

[30:34] Nikola Tsarigradski: in chess, in the, in the Elo system of FIDE, the governing chess body,

[30:39] Nikola Tsarigradski: there’s this rule that the rating you gain from beating a weaker player is capped.

[30:46] Nikola Tsarigradski: So no matter how big the rating difference is, you always get 0. 8 rating points for A win.

[30:54] Nikola Tsarigradski: Even if the model would say you should actually get 0. 01, you always get 0. 8.

[31:02] Nikola Tsarigradski: And the reason is exactly this, that if you play against much, much weaker players,

[31:07] Nikola Tsarigradski: the model can’t accurately measure win rates over a small sample size.

[31:14] Nikola Tsarigradski: Like whether your win rate is like 99. 9% or 99% is for the model, a huge difference. This might be like a difference of like 600 Elo points,

[31:24] Nikola Tsarigradski: but if you play five games, you’re just going to win all five games. So there’s no way to infer anything on the game result.

[31:29] Nikola Tsarigradski: And the reason this has been a drama in the chess world lately is because like most famously current number two in the chess world, Hikaru, probably heard of him,

[31:39] Nikola Tsarigradski: is currently playing a lot of low rated tournaments. So he’s like going to a lot of state championships where he’s playing against players who are like a thousand points weaker than him because he needs to reach a certain number of minimum number of games played in a year to be eligible to compete in the world Championship or like the candidates tournament.

[31:59] Nikola Tsarigradski: Like he has to just like play a certain number of games and not lose rating. So he’s going to a lot of these like really, really low rated tournaments, which doesn’t really make sense, but he just needs to like get a certain number of games.

[32:11] Nikola Tsarigradski: But now he gets basically like a free rating point every single time he beats someone who has zero chance of eating him.

[32:19] Nikola Tsarigradski: So now basically if he kept doing this, he could reach like an all time high chess rating just by beating up weaker players, which kind of like defeats the system.

[32:32] Nikola Tsarigradski: I’m pretty sure as far as understand it’s like not his goal with this, but it’s kind of like a, a side effect of the rating system having like this cap 0. 8 for each one you get.

[32:41] Travis: I imagine it’s to some degree looked down upon or do you have any sense?

[32:47] Nikola Tsarigradski: Yeah, sure. Like it’s just an exploit.

[32:49] Nikola Tsarigradski: It’s fundamentally like an issue of the rating system. I don’t think it’s appropriate to criticize a player for abusing this.

[32:57] Nikola Tsarigradski: If you are abusing it, really gain rating and then like gain advantage from an increased rating,

[33:03] Nikola Tsarigradski: that would be a problem.

[33:06] Nikola Tsarigradski: But ratings are self correcting. So if you do this and you increase your rank artificially,

[33:12] Nikola Tsarigradski: as soon as you start playing against players of your own skill, your rating will fall back again.

[33:19] Nikola Tsarigradski: But I don’t know, it’s just like a quirk operating system where, I don’t know, it suddenly shows that there’s an issue with the system that people did not foresee.

[33:28] Nikola Tsarigradski: I’m pretty sure that FIDE came up with a way to fix it. I’m not sure what they did, but I read some news on this that FIDE did come up with some solution to this to basically not allow it to just farm rating.

[33:41] Travis: And then when you think about tying that back to go,

[33:44] Travis: we don’t have such an issue correct? I mean, because we’re using each federation or each group’s dan-kyu system,

[33:54] Travis: is it at all linked or tied back to any of these rating systems?

[33:58] Nikola Tsarigradski: So I think the main point of this conversation is how this ties back with the dan-kyu system,

[34:04] Nikola Tsarigradski: because this actually causes a major conflict and the different Go federations have different ways of trying to to resolve this.

[34:11] Nikola Tsarigradski: It’s not really obvious, but the main issue with trying to unify the Elo system with the dan and kyu system is that the dan and kyu system is predictive in and of itself,

[34:26] Nikola Tsarigradski: but not for a win rate. It is predictive for handicap.

[34:31] Nikola Tsarigradski: Like this is the nature of the dan-kyu system.

[34:34] Nikola Tsarigradski: A difference in rank predicts a number of handicap stones that should lead to a fair game.

[34:41] Nikola Tsarigradski: This is basically how the, how the system is defined. Whereas the Elo system is defined by saying, okay,

[34:48] Nikola Tsarigradski: 100 Elo difference is 63% win rate regardless of skill level.

[34:54] Travis: Got it. So the, the point of focus and the goal is theoretically different.

[34:59] Gazza White: Yes.

[34:59] Nikola Tsarigradski: And the problem is, and a lot of people I talk with about this intuitively get this wrong, which makes a lot of sense,

[35:07] Nikola Tsarigradski: it’s that not every Go rank is made equal with respect to Elo.

[35:16] Nikola Tsarigradski: So if you are a 12 kyu,

[35:21] Nikola Tsarigradski: a 10 kyu might give you two stones and you have a fair game.

[35:24] Nikola Tsarigradski: If you are 1 dan, a 3 dan might give you two stones and you have a fair game.

[35:28] Nikola Tsarigradski: But if you play the game without any handicap,

[35:33] Nikola Tsarigradski: the 12 kyu will have a significantly different win rate against the 10 kyu than the 1 dan will have against the 3 dan.

[35:41] Nikola Tsarigradski: Because the stronger you get at go,

[35:44] Nikola Tsarigradski: the higher the numerical value of a handicap stone becomes.

[35:49] Nikola Tsarigradski: I guess the intuitive way to think about it is that a stronger player can handle a handicap stone better.

[35:57] Nikola Tsarigradski: Or the other way to think about it is to say that the stronger you get,

[36:00] Nikola Tsarigradski: the more additional skill and knowledge you need to actually get another handicap stone worth of rank.

[36:06] Travis: Is that the same as saying if there’s one stone added to the board of handicap, it’s that much more difficult to overcome it because. Because we’re at higher ranks, or am

[36:16] Nikola Tsarigradski: I thinking of it in the wrong Way,

[36:18] Nikola Tsarigradski: not necessarily. Like, it’s not. It’s not about. Like, it’s not about the actual handicap stones. It’s about arguing about the win rate in games without any handicap, with.

[36:27] Travis: Okay, got it.

[36:28] Nikola Tsarigradski: Basing the matchups on. On handicap. There are statistics on this. Like there’s a page on Sensei’s Library which is just called EGF rating system. And there’s a pretty good explanation of this with respect to the EGF rating system.

[36:43] Nikola Tsarigradski: You can basically look up the predicted win rates of a player against a stronger player.

[36:52] Nikola Tsarigradski: There’s this table that tells you for each given rank,

[36:57] Nikola Tsarigradski: what the predicted win probability is against the stronger player.

[37:01] Nikola Tsarigradski: And the system itself predicts that, for example, a 20 kyu will beat a 19 kyu 38% of the time,

[37:10] Nikola Tsarigradski: whereas a 6 dan will beat a 7 dan only 20% of the time.

[37:15] Travis: Wow, that’s a. Yeah, that’s a difference. So you’re saying. Yeah, that change in win rate, there’s a constant change in win rate between the differences as you get to different levels of strength.

[37:25] Nikola Tsarigradski: Yes.

[37:26] Nikola Tsarigradski: And what this means is just that if you want to unify Elo and the dan-kyu system,

[37:31] Nikola Tsarigradski: one stone handicap is not always the same number of Elo points.

[37:37] Nikola Tsarigradski: The difference between a 6 and a 7 might be 150 Elo points, whereas the difference between 9 kyu and 10 kyu might only be 50.

[37:44] Travis: Got it.

[37:46] Nikola Tsarigradski: And there are actually statistics on this. So you can basically convert the Elo system into a dan-kyu level system.

[37:57] Nikola Tsarigradski: Statistics show us that this is pretty probably logarithmic transformation.

[38:04] Nikola Tsarigradski: So the Elo points difference between a single rank is pretty similar for the DDK ranks. And then the stronger you get,

[38:14] Nikola Tsarigradski: the jumps in additional Elo you need to get a full handicap stone worth of rank increases basically exponentially.

[38:24] Travis: I see. Okay, got it.

[38:26] Travis: And this is really probably born out of the fact that the dan-kyu system,

[38:34] Travis: being it based on, you know, a stone, a physical stone,

[38:37] Travis: was probably likely quite limited because of its sort of beginnings. Right. They didn’t have this level of statistical inference and ability hundreds or thousands of years ago. And this was a simple way to do it, correct?

[38:50] Nikola Tsarigradski: Yes, exactly, exactly.

[38:53] Nikola Tsarigradski: And I think it’s really, really elegant.

[38:55] Nikola Tsarigradski: It just means that we need to be really, really careful in translating Elo to orangs. And the EGF has a way of doing this, and the American association has a way of doing this.

[39:07] Nikola Tsarigradski: And the way the American association does it, I think is almost like criminal.

[39:12] Nikola Tsarigradski: They’re basically like an Elo rating where like 100 points are one rank difference.

[39:19] Nikola Tsarigradski: And the difference in rank does not matter for like everything is 100 elites. Basically the American possession just ignores this problem and just accepts that the ratings within the really, really strong players properly change too fast.

[39:38] Nikola Tsarigradski: And the ratings for the really, really weak players change too slow, slowly. And they also ignore that if you then have big differences in rank and you play a game without handicap,

[39:50] Nikola Tsarigradski: then the win rate prediction of the system is probably wrong.

[39:55] Travis: It’s linear, right? It’s simply like 100, 100. Got it. It’s not logarithmic like you’re saying if

[40:00] Nikola Tsarigradski: a 3 dan plays a 5 kyu, then the, the Elo system of the American association will predict a win rate. But this win rate will probably not be accurate to true win rate.

[40:10] Nikola Tsarigradski: It’s not a huge problem because these matchups don’t happen that often. And if they do,

[40:15] Nikola Tsarigradski: the win rates are so close to one that it’s not a huge problem.

[40:20] Nikola Tsarigradski: But mathematically it’s far off. It’s really, really far off.

[40:23] Travis: Interesting.

[40:24] Nikola Tsarigradski: And the, the egf,

[40:26] Nikola Tsarigradski: they basically do a logarithmic transform of an Elo into the EGF rating system.

[40:33] Nikola Tsarigradski: So within the calculation of the EGF rating,

[40:37] Nikola Tsarigradski: we are basically calculating an Elo and then, then pressing it through a logarithmic formula to normalize it into a rating where 100 points equate one stone difference.

[40:50] Nikola Tsarigradski: But within the EGF system, if a 5 dan beats a 5 dan,

[40:54] Nikola Tsarigradski: they might only get four points. If a 15 kyu beats a 15 kyu, they might get 20 points.

[41:00] Travis: Okay,

[41:02] Travis: and Nikola, how does this tie into a question I had earlier around using a sort of strongest player or strongest rating point of reference in a rating system?

[41:16] Nikola Tsarigradski: Go is extremely beautiful in that it does not need that.

[41:22] Nikola Tsarigradski: So with chesses is a problem we can never say is is Magnus really better than Fischer? But hard to say.

[41:29] Nikola Tsarigradski: Like rating inflation and deflation is really hard to take into account.

[41:33] Nikola Tsarigradski: Best thing you can do is maybe train AI models to evaluate the level of play and say, ah, probably this play looks better.

[41:40] Nikola Tsarigradski: But we covered that the win probabilities of two players who are one rank apart differ depending on the rank. So like, the stronger you get, the higher the win probability is for the stronger player.

[41:58] Nikola Tsarigradski: This means that if you identify players that are exactly one rank, right, they play like no Komi games and realize, okay, no Komi games, they have 50% win rates in games with Komi, the stronger player wins some more games.

[42:14] Nikola Tsarigradski: You can define 1 dan as being the exact strength where a player has a certain win rate against a player who is one Rank weaker.

[42:28] Nikola Tsarigradski: So there is an objective measure of what 1 dan actually is.

[42:32] Nikola Tsarigradski: And the EGF just like happens to define this as 27%.

[42:41] Nikola Tsarigradski: So a player who loses against a one stone weaker player 27% of the time is defined as 1 dan.

[42:51] Nikola Tsarigradski: And this is pretty cool. I mean we do not really follow this that much.

[42:59] Nikola Tsarigradski: In theory, this would change over time. Right. Like when we let the system play out,

[43:04] Nikola Tsarigradski: where this 27 win rate against a one-stone-weaker player actually is, might shift. Like it might become that. Now a 3 kyu has this win rate and the win rate for 1: 1 is actually better.

[43:18] Nikola Tsarigradski: The system does not self correct for this. But what we in theory could do is we could look at how the database has shifted over time and we could correct for it.

[43:28] Nikola Tsarigradski: So if we, if we realize that the player who has this exact win rate is not the one that anymore, but it shifted a little bit, we could correct it by just like shifting everyone in the database a little bit again to match this number.

[43:41] Nikola Tsarigradski: And I think this is really, really beautiful. Like chess doesn’t have anything.

[43:44] Nikola Tsarigradski: Yeah, it’s difficult to measure. That’s of course a downside because you need like a huge database of games being played between people who are exactly one ranked part to be able to accurately tell it.

[43:56] Nikola Tsarigradski: Okay, it’s actually 27%. It’s not 28, it’s not 25, it’s 27.

[44:00] Nikola Tsarigradski: So you need a large database in order to make this assessment.

[44:06] Travis: Again, I’m more familiar with the OGS system, but like we see that they’ve made some adjustments two or three years ago. They,

[44:13] Travis: you saw a sort of. I think it was a downward shift in the overall ratings of people across the board.

[44:19] Travis: Is there some degree to which that’s what’s happening they’re trying to do in adjusting to what the data set and the population is doing?

[44:28] Nikola Tsarigradski: Generally speaking,

[44:30] Nikola Tsarigradski: yes. Yeah, you want to try to keep the ranks stable.

[44:36] Nikola Tsarigradski: The way this depends on the population is that if you have like a population that does not improve but gets a lot of new players,

[44:44] Nikola Tsarigradski: then the ratings will inflate because basically everyone will like steal the rating points of the new players without actually getting better. But then on the other hand, if we don’t get any new players,

[44:53] Nikola Tsarigradski: everyone improves, then the ratings will get deflated.

[44:57] Nikola Tsarigradski: So basically like you’re. I don’t know, let’s say you’re 1 dan.

[45:00] Nikola Tsarigradski: You take half a year to study. You practice a lot, you get better, you get like a real stone better.

[45:06] Nikola Tsarigradski: But assume that everyone else did exactly the same. Everyone Got a full stone better. You will go to next tournament as a 1 dan. And even though you’re playing at a 2 dan level now,

[45:14] Nikola Tsarigradski: everyone is playing rank stronger as well. And then the entire rating system just like deflated by one full rank because you can’t get new rating because everyone else does it.

[45:23] Nikola Tsarigradski: Okay,

[45:24] Nikola Tsarigradski: so you always try to like find the balance between the growth of your population of new players and the improvement rate of your population.

[45:33] Nikola Tsarigradski: And depending on how many new people get to go or how much people invest into studying, this may change over time. I’m pretty sure that the EGF has a rating commission that looks into this every few years and they have a bunch of ways to control for this.

[45:48] Nikola Tsarigradski: They’re basically there to fix this inflation and deflation. So there’s like a rating commission that looks at the data of the past few years and if the ratings tend to get inflated,

[45:57] Nikola Tsarigradski: we just add like a negative bonus term. So basically everyone loses like a flat amount of points for every game they play.

[46:04] Nikola Tsarigradski: If you play two games against same level opponent and you win one, you lose one, you lose a tiny bit of rating.

[46:11] Nikola Tsarigradski: Currently the system is deflating, so we have a positive bonus term.

[46:15] Nikola Tsarigradski: So right now you get basically like a flat amount of free rating for every game you play.

[46:21] Nikola Tsarigradski: And this is just to combat inflation and deflation.

[46:24] Nikola Tsarigradski: There is one thing I still would like to touch upon quickly please.

[46:28] Travis: Yeah,

[46:30] Nikola Tsarigradski: generally speaking, like we want our rating system to tell us basically who’s better than someone else. And we want them to give some level of prediction about how games will go.

[46:41] Nikola Tsarigradski: And there is one specific thing which rating systems can’t account for, which I find pretty funny. And I’ve been wondering for a long time if there’s like a rating system could account for that.

[46:51] Nikola Tsarigradski: In a sense, the rating systems always assume that every player is the same. If two players have the same rating, they’re like exactly equally skilled and they have the same win rate against everyone else.

[47:02] Nikola Tsarigradski: But that’s clearly not the case. Right? You can have three players who are like exactly the same level and each one will have a positive win rate against one of the others just because they’re different human beings play matches or mismatches.

[47:14] Nikola Tsarigradski: And this is something I, I looked up a while ago. There are like within top level Go in Europe there are a bunch of these like I call it love triangles of people who have basically a non transitive win rate against one another.

[47:28] Nikola Tsarigradski: There is the European pro, Artem.

[47:31] Nikola Tsarigradski: He has a win rate of 9 to 1 against the European 6 dan player Viktor Lin.

[47:39] Travis: Okay.

[47:40] Nikola Tsarigradski: That guy in turn has a 23 to 15 win rate against the European 7 dan Lukáš Podpěra.

[47:47] Nikola Tsarigradski: And Lukáš Podpěra in turn has 10 to 7 win rate against Viktor Lin. So they all have like clearly positive win rates against one another. Like everyone is better than the next.

[47:58] Nikola Tsarigradski: This clearly means that like the ratings you have in an Elo system or like even with Go rating system,

[48:03] Nikola Tsarigradski: they predict your win likelihood against like a general player of the probability,

[48:08] Nikola Tsarigradski: general player of the population.

[48:10] Nikola Tsarigradski: But the probability does not necessarily match for any individual.

[48:14] Nikola Tsarigradski: It might be that against any individual I have a higher or lower performance.

[48:19] Nikola Tsarigradski: I’ve been wondering if there’s a way to account for this.

[48:23] Nikola Tsarigradski: Maybe this whole history rating system that’s used for the goratings.org Might for example,

[48:31] Nikola Tsarigradski: take into account within the model what your specific performance has been against an individual player.

[48:37] Travis: Hmm.

[48:38] Nikola Tsarigradski: But I, I don’t know if there’s like any way to do it efficiently. Especially I don’t know if there’s any way to do it without breaking some of the things that you actually want to have.

[48:45] Nikola Tsarigradski: Hmm.

[48:46] Nikola Tsarigradski: That’s always the issue. Like sometimes you, you do something with like breaks something else that you won’t have.

[48:50] Travis: Interesting. Yeah, fascinating. What else should we know about rating systems before we go?

[49:00] Nikola Tsarigradski: Um,

[49:01] Nikola Tsarigradski: one of the takeaways from rating system like this is that they’re basically applicable to everything.

[49:09] Nikola Tsarigradski: Like chess was kind of the pioneer of developing predictive and self correcting rating systems. But you can use them in absolutely everything.

[49:17] Nikola Tsarigradski: And people do,

[49:19] Nikola Tsarigradski: for example, like the ratings of the international soccer teams based on an Elo system.

[49:27] Nikola Tsarigradski: Anything where you have like a one on one,

[49:29] Nikola Tsarigradski: you can use these self correcting systems, usually like a Glicko system,

[49:33] Nikola Tsarigradski: to predict performances. So in that sense chess and Go have, I mean mostly chess, I guess has been really helpful to many other domains with go. It’s just a problem that it kind of conflicts with the dan-kyu system.

[49:47] Nikola Tsarigradski: So we already have a system in place that does something else really, really well,

[49:51] Nikola Tsarigradski: which is like allow players of very different skill to have a meaningful game.

[49:55] Nikola Tsarigradski: That’s pretty much impossible in chess, right. Where a grandmaster can’t really have a meaningful game against a 1500.

[50:02] Nikola Tsarigradski: But it comes kind of at the cost that it breaks some of the properties that we want to have from the ecosystem.

[50:09] Travis: What do you think would be some positive changes in a good direction in terms of how we use these rating systems or any sort of converging of a rating system in Go?

[50:22] Nikola Tsarigradski: I guess this loops a little, a little back to this idea that like a rating can Never be perfect and it makes only sense to optimize it according to your needs.

[50:32] Nikola Tsarigradski: And it makes sense to identify any problems or rating system. For example, like one, one problem that we didn’t touch upon is rating resets. This is like a, kind of like a big topic here in Finland because there’s not that many tournaments.

[50:48] Nikola Tsarigradski: People improve and then if you just let them play a tournament and gain their true rating by playing, they will just like steal everyone else’s rating.

[50:59] Nikola Tsarigradski: Then it will take like a long time until their rating actually matches.

[51:03] Nikola Tsarigradski: So this is something that has been being discussed here quite a lot of like how do we reset ratings or how do we deal with players who very, very quickly change their rating without affecting the rest of the population.

[51:15] Nikola Tsarigradski: A lot of I’ve mostly looked at the EGF system. That’s the one I’m most familiar with and I do consider it to be quite good.

[51:28] Nikola Tsarigradski: I think the most important thing is that within the Federation there should be people with expertise who are responsible for looking whether the system performs up to what we expect from it and who also identify issues.

[51:47] Nikola Tsarigradski: Like for example the issue I mentioned earlier with the FIDE ratings and Hikaru Gate.

[51:53] Nikola Tsarigradski: Like when. When stuff like this happens, there has to be some other expertise in the Federation to recognize it and then like propose a change or a fix.

[52:00] Nikola Tsarigradski: Personally, I currently don’t have any grievances with the EGF system. I think it works pretty well.

[52:07] Nikola Tsarigradski: Got it. Excellent.

[52:08] Travis: Well Nikola, this has been really helpful. Thanks for your time.

[52:11] Nikola Tsarigradski: I’m so glad you let me be on the on the part.

Interview with Gazza White on the Sopalkosanol World’s Top Player Championship

[52:32] Travis: We’re going to turn to a different topic and go back to talking with Gazza White about out some new professional tournaments that have come onto the scene over the past year.

[52:44] Travis: And so let’s jump to that conversation now.

[52:50] Gazza White: It’s actually been quite an influx of new international tournaments that have come out.

[52:57] Gazza White: Let’s see, Korea expanded one of their domestic competitions to a biannual so one every two years.

[53:06] Gazza White: International composition.

[53:07] Travis: Okay.

[53:08] Gazza White: Korea had a. Well, Korea they still do. They have a competition called Korea’s Strongest where a league of nine players will play a round robin and the winner of that league will play the current champion in a best of five.

[53:29] Gazza White: So they held that every year from 2000 to 2024 Shin Jin-seo won all five occasions and then this year they expanded it to an international version. So World’s Strongest Player and which nine top players.

[53:45] Gazza White: So I think,

[53:46] Gazza White: let me just confirm.

[53:50] Gazza White: Yes, four from Korea, three from China, one from Japan and one from Taiwan.

[53:55] Gazza White: Okay, so they Participated in a round robin. So nine players, so eight games each. And the top two played a best-of-three final and the top two were Tu Xiaoyu and Shin Jin-seo.

[54:09] Gazza White: So they played each other and Shin Jin-seo won two games to one.

[54:14] Travis: Wait, was this the one where Shin Jin-seo, whenever he won the first game in the finals match, he would win that final but if he loses the first game in a best of match for final round, then he like doesn’t win.

[54:30] Travis: Am I making this up?

[54:31] Gazza White: You’re close. What we were saying is that if he lost any game in the final match,

[54:36] Gazza White: then he would end up losing the final Only in international tournaments. In domestic tournaments, that.

[54:43] Travis: That’s. That one hold up.

[54:45] Gazza White: I see in international tournaments, every tournament that he won, he won with a perfect record.

[54:51] Travis: Wow.

[54:52] Gazza White: But in this case, he actually.

[54:54] Gazza White: Well, in the round robin, he lost two of his games. He lost to Xu Jiayang and Tu Xiaoyu. So he won six and lost two.

[55:02] Gazza White: And then in the final he lost the first game to Tu Xiaoyu. But then he ended up winning the second and third games.

[55:10] Gazza White: So yeah, he.

[55:12] Gazza White: He overcame the hurdle of losing a game in the tournament and ended up winning this one. It was a very interesting tournament. The scheduling was quite staggered. They had some games in March, some games in June,

[55:29] Gazza White: and then the final was in September.

[55:32] Travis: Gazza, why did they do this? Why did they stagger out games over such a long period of time sometimes versus saying, hey, we’re going to do this in a shorter period.

[55:41] Travis: Is it logistics? Is it how tiring it is on the players schedules? Do you have any sense?

[55:49] Gazza White: I think it’s mainly logistical. The domestic version, like I said the previous five years, it had a domestic version and they would have eight rounds because it’s a nine player tournament.

[56:01] Gazza White: So each player would play eight games.

[56:05] Gazza White: So they would have.

[56:06] Gazza White: With a buy, they would have nine rounds and each round would be over a week.

[56:11] Gazza White: Each week they would have one of those rounds. And I think mainly it’s for TV purposes because they broadcast these on the TV channel in Korea.

[56:22] Gazza White: So to have it over nine weeks maintains that engagement with the sponsors and things like that.

[56:29] Gazza White: But I think the reason why they had this sort of in two separate periods in March and then in June.

[56:37] Gazza White: Again, logistically,

[56:39] Gazza White: the Chinese players can’t just keep flying over every week,

[56:43] Gazza White: for instance.

[56:44] Gazza White: So they would have them come over, play three or four games in, you know, one visit,

[56:50] Gazza White: go back,

[56:51] Gazza White: and then three months later they would come over,

[56:55] Gazza White: play the rest of their games and then go back.

[56:57] Travis: Got it?

[56:58] Gazza White: You also have to schedule it in between the other international players commitments,

[57:03] Gazza White: like all of their domestic tournaments and any other international tournaments that are happening.

[57:09] Gazza White: Obviously, the other alternative would be to just have them come over, play all eight games over,

[57:15] Gazza White: you know, maybe two weeks or something. But I think.

[57:18] Gazza White: I imagine it would be just about impossible to find two whole free weeks for all the international players to be available.

[57:28] Gazza White: So I imagine all schedulings have their advantages and disadvantages.

[57:33] Travis: How did they even start a tournament like this? If they say, okay, it’s been a domestic one, we’re going to make it an international one, hopefully get a very strong field, which it sounds like you’re saying they did.

[57:47] Travis: How are they selecting the players from outside?

[57:51] Travis: Do you have a sense? Is it really purely invitation only? Is it based on ranking? Is it all of the above?

[57:57] Travis: Is it an open tournament where they’re saying, hey, anybody can. Any professional can try.

[58:02] Travis: How is that working?

[58:04] Gazza White: Specifically for this tournament,

[58:07] Gazza White: the Chinese, Japanese and Taiwanese players were chosen by their respective federations.

[58:12] Gazza White: So essentially it was up to those respective federations to select a player. So I. I would imagine that they said, China, you get three spots. Japan, you get one spot.

[58:23] Gazza White: Taiwan, you get one spot. And then left it up to the federations to determine the selections. For instance, China had a preliminary tournament,

[58:35] Gazza White: I believe, actually China was offered only two spots and then there was a wild card. The wild card was Ke Jie.

[58:44] Gazza White: In fact,

[58:45] Gazza White: he declined.

[58:47] Travis: Whoa.

[58:49] Gazza White: Again, it would have to be because of the LG Cup.

[58:52] Gazza White: And he was replaced by Dang Yifei,

[58:57] Gazza White: who was also a Chinese player. And I think Daniel was the number one Chinese player at the time,

[59:03] Gazza White: so he was the replacement.

[59:06] Gazza White: I’m not really sure how Japan selects it, but I imagine that they have players in their national team and they may have had them play against each other, or they may have just selected someone based on availability.

[59:18] Gazza White: But they selected Fukuoka Kotaro,

[59:21] Gazza White: who was a fairly young player,

[59:24] Gazza White: but he’s in the national team.

[59:26] Gazza White: And Taiwan,

[59:27] Gazza White: their selection was Hsu Haohong, who has been by far the strongest player in Taiwan for quite a while now. I’m not sure exactly how he was selected, if he was just picked because he’s the best or if they went through qualifiers.

[59:39] Travis: But they’re like, you beat Shin Jin-seo you’re going. I don’t know, like, right. Isn’t. Didn’t he.

[59:46] Gazza White: He did beat Shin Jin-seo. So at the Asian Games a couple of years ago,

[59:50] Gazza White: yes, he was the first Taiwanese to beat Shin Jin-seo, but Lai Junfu actually also beat Shin Jin-seo. So last year in the Kuksu Mountains International.

[59:59] Travis: That’s right.

[59:59] Nikola Tsarigradski: Oh wow.

World’s Strongest Player Selection and Closing Sponsors

[01:00:00] Gazza White: There are two now.

[01:00:02] Travis: So it sounds a bit like the Asian Games or like an Olympic Games. It’s right. The federations,

[01:00:08] Travis: they say, hey, it’s up to you to decide which players do you want to bring. And then each of those federations, as you articulated, decide how to determine that. They look at schedules, they look at strongest players.

[01:00:19] Travis: Maybe they have them duke it out in some private matches. Interesting.

[01:00:23] Gazza White: Yeah. With the Korean team, I believe Shin Jin-seo was automatically.

[01:00:28] Gazza White: He was automatically qualified as the current domestic Korea’s Strongest Player champion. And then they had preliminary tournaments for the other three participants.

[01:00:39] Travis: Got it?

[01:00:39] Gazza White: Which were Park Junghwan, Kang Dong-yun and Shin Min-jun.

[01:00:44] Gazza White: That’s the international world’s strongest player.

[01:00:49] Gazza White: I believe that it will go back to a domestic version of Korea’s Strongest Player. And then in 2027,

[01:00:56] Gazza White: it’ll be world’s strongest player again. So it alternates.

[01:01:00] Gazza White: That was the arrangement that I heard. It alternates in even years,

[01:01:03] Gazza White: it’s just the domestic version and odd years international version.

[01:01:08] Travis: Very interesting.

[01:01:13] Travis: That wraps up our episode. I hope you enjoyed it and I will catch you next time. Thanks so much.

[01:01:31] Travis: Foreign I want to give one final thank you to our sponsors,

[01:01:38] Travis: Go Magic, who creates interactive courses, Go problems, lessons and lectures on the Game of Go.

[01:01:44] Travis: Everything you need to know to learn the Game of Go online@go magic.org AI Sensei where you can play Go against the strong KataGo AI. Analyze your uploaded games,

[01:01:55] Travis: discuss moves with the community and turn mistakes into Go problems.

[01:02:00] Travis: Also, SmartGo One, your complete app for the game of Go.

[01:02:03] Travis: Learn to play, practice against the computer, study master games, solve problems and read Go books. Free to download in your app store.

[01:02:12] Travis: And finally, Baduk Pop. Learn the rules of the Game of Go with a fun interactive tutorial. Sharpen your Go skills with daily random Go problems at your choice of difficulty.

[01:02:22] Travis: Play games online or with a variety of AI opponents you each with its own unique playing style and strength. @BadukPop.com that’s B A D U K P O P dot com.