Glicko-2 does not require accounts to start at 12 kyu. So why do we do that?

This topic is in response to yet another new topic in which a beginner is complaining about getting trashed by much stronger opponents and the OGS advice is to just take a few beatings to establish a rank. :pensive:

OGS uses the Glicko-2 rating system. This has advantages of stability, speed of response, etc. However, as has been noted multiple times on this forum, this has the restriction that new accounts start at 12 kyu, which in turn leads to the situation of beginners getting matched against much stronger players until their rank is established.

But this is actually not true! :grin:

Look what the Glicko-2 algorithm has to say about starting values:

So the new accounts on 12 kyu (rating of 1500) comes from that. Except new accounts are not necessarily unrated! If someone is an absolute beginner, they are ranked 30 kyu (and have the corresponding rating). If some one knows their rank from another system, then they have a rating. The only players that are truly unrated are those who are not beginners but who also don’t know their level (maybe they have been playing casual games IRL with unranked players for a while but never played within any rated system). Therefore, it should be possible to allow new accounts to declare their rank (if known), declare themselves as a beginner (i.e. 30 kyu) or just accept the default (i.e. 12 kyu, no change from current system). Deviation and volatility can just take default initial values. Any incorrect declaration should sort it self out within a few games, same as current system. Difference is, at least beginners (and strong players too actually) will have fairer match ups during initial period while deviation is reduced. :sunglasses:

I can hear the objections already. “What about the sandbaggers!?” Well, the current system is hardly immune to sandbagging, is it? So at least this would allow honest players to get fair match ups straight away. Alt sandbaggers could declare 30 kyu and pick on beginners. But their rank will go up within a few games so they are limited to that just as they are now so not really any different from the current system. As a protection from getting sandbagged, you could allow players with established rank to reject games from players with high deviation (i.e. [?] accounts) - could be useful for players who remain at low kyu level for a while and keep getting sandbagged and are fed up with it - but that is a separate issue Re sandbagging, as opposed to this topic’s subject of initial rating value. :slightly_smiling_face:

Tagging @anoek as he knows the inner workings of the rating system. What do you think about declared rank being used for initial rating? Or have I missed something? Cheers :beer:

4 Likes

If match making is the problem, why fiddle with the ranking system? Glicko has no restrictions for match making.

5 Likes

I understand wanting to restrict handicap games to 9 stones, but agree even games could be opened up to anyone vs anyone (as long as you can still set rank ranges for challenges)

1 Like

Handicap doesn’t even work as you expect when a player has a provisional rank. Automatic handicap is drastically reduced in this case.

1 Like

4 Likes

<Insert standard comment #6>

2 Likes

Topic title edited as the original looked more like a support request.

The replies so far have not addressed the main point so I will summarise here:

Glicko-2 does not require new accounts to start on a rating of 1500. Doing so causes problems. So why do it? Why not start accounts at a best guess rank, e.g. rank from another server / association if known or 30 kyu for beginners?

Because match making is the purpose of the rank system so incorrect match making due to incorrect initial rank is a sign that the rank system has a problem. The problem is incorrect initial rank. The argument against declared rank is that Glicko-2 doesn’t support it. But that is incorrect - see quote above. Arguments against it Re sandbaggers are also invalid. Can anyone give a good reason why not to have initial rank set by user declaration?

I don’t understand. Why should it be restricted? Why not have a complete beginner Vs a SDK with 20+ handicap? Could make an interesting game. Might not be everyone’s cup of tea but why restrict?

I agree but not just even games. One can assign a theoretical probability of win for any game based on player ratings, board size, komi and handicap. The bigger the probability in an unexpected result, the larger the rating change. The bigger the probability in an expected result, the smaller the rating change. Then any game can be rated. This is essentially what systems like Glicko-2 are doing with their algorithm. Indeed, Glicko-2 says nothing about restricting rating differences. So again, why restrict?

Further evidence that default initial rank rather than declared initial rank is a problem.

If you take a closer look at b) it states you should use the players most recent rating, RD, and sigma, which only exist in the context of Glicko2.


Also the 1500 is an arbitrary number. You can reduce the rating of all player by 1000 and nothing relevant would change. If we would have choosen 500 as the initial rating, we would have the same problem and discussion. This makes we wonder what the impact of setting individual ratings on the overallsstability is.

2 Likes

I agree that 1500 is arbitrary :+1: The fact that 1500 corresponds to 6 kyu basically says that if everyone is given the same starting point of 1500 and Glicko is allowed to do it’s thing, then the players that end up with this arbitrary rating of 1500 are those of a strength of approximately 6 kyu in other systems (AGA, EGF, etc.), as determined by surveying members for anchor points. Therefore, we will call this arbitrary rating of 1500 by a certain rank to give it real world meaning and that happens to be 6 kyu. Choosing a different starting value like 1000 or 500 will just change the conversion between internal rating and external rank, e.g. 1000 = 6 kyu or 500 = 6 kyu. So, yeah, that would be pointless. :laughing:

But I’m not saying to change the starting value of all players by the same amount to some new default like 1000 or 500. I’m saying, now that the system is established with some anchoring to other systems, new accounts can join at the approximately correct level if known (i.e. rank from another system if known or 30 kyu for raw beginners) so new accounts will join at a variety of levels that are closer to where they are likely to end up under the current system of starting them at the likely incorrect value of 6 kyu. Deviation and volatility would still be unknown so just use the default values.

Looking at the Glicko equations, this will work fine. If anything, it will improve stability. Essentially, the Glicko equations mean that:

  • Playing stronger opponents has more potential to increase your rating than playing weaker opponents (step 3, definition of E ~ 1 / (1 + exp(mu_j)) where mu_j is opponent rating). The size of this effect is relative to your current rating (the above is actually E ~ 1 / (1 + exp(-(mu - mu_j))) where mu is player rating). Therefore, having an incorrect initial rating means you will adjust faster. With an approximately correct initial rating, you are already in about the right place so there’s less adjustment to do anyway.

  • Opponents with a high deviation are weighted less favourably than opponents with a low deviation (step 3, definition of g ~ 1 / sqrt(1 + phi^2) where phi is deviation). Therefore, provisional accounts (high deviation) do not affect their opponents’ ratings as much as games with established players. Hence, the impact of incorrect initial rating is relatively small on the opponents’ ratings compared to, say, sandbaggers with non-provisional rank who decrease their rank on purpose. However, allowing declared rank and thus approximately correct initial rank would reduce the effect on the rank of the provisional player’s opponents even more. So, in effect, there is an extra element of rating system instability by having usually incorrect initial rating that would be eliminated by having declared initial rank and thus approximately correct initial rating.

  • Winning will increase your rating while losing will decrease your rating (step 4, definition of Delta). Again, results against opponents with a high deviation are weighted less favourably. This aspect of the equations doesn’t make too much different to the current discussion. The important points are the two above.

So, actually, when you consider the above, declared rank is better for new accounts (especially beginners) and better for established accounts as well (both due to better matchmaking) and it’s also better for the rating system stability. So, there is no reason not to (apart from time and effort to implement) and lots of positive reasons to do it. :slightly_smiling_face:

2 Likes

Maybe someone will say it has been determined that we need beginners coming in around 12k to minimize drift of the system over time. If lots of people come in at 30k, rank up to 15k, then quit, they are essentially sucking rating points out of the pool.

But if that’s true, we really just need the average to be around 12k, and that can be controlled even while giving incoming new players some flexibility - just shift the assigned values up or down a bit (or a lot) as necessary to achieve whatever average we need. Maybe we can even let people choose their exact rank but bias the initial uncertainty to produce the same effect?

Also, there are other ways to inject rating points than using the poor beginners, for example by fixing the rank of some popular bots or just directly inflating everyone’s rank over time.

1 Like

Nope, because rating changes are not a zero sum change, the new 30k? has low confidence so could gain 100 points from beating a solid 28k who only loses 10.

The only valid argument I’ve heard for why new users can’t choose their rating or default to 30k is it makes life easier for sandbaggers. But the cure for that punishes real beginners with games against too strong opponents so I’d rather make joining OGS a more pleasant experience for real beginners and find another way to deal with the sandbaggers (I assume the former outnumber the latter).

4 Likes

In my eyes the only two fair options are 1. Everyone starts in the middle and 2. Everyone can pick their starting rank.

If we make dans play their way up from 30k they’re gonna need a lot more games to get an accurate rank and all the kyus will hate even more playing ? accounts because they’ll be destroying them on their longer climb up.

1 Like

Again? OMG.

Not true any more. I will quote myself:

"Before the recent rank adjustment, players started at a psudo-12k and were matched by humble rank (which GreenAsJade described in that other thread). The recent rank adjustment seems to have broken humble rank, and the new starting pseudo-rank is 6k. Anyone can verify this by watching some games in the overview thumbnails (the only place so far as I know where the pseudo-ranks are actually displayed). You will see, for example, a 6k playing, and upon checking his profile find that this is the first game of a provisional (?) player. I have done this. Moreover, on September 6 I created an account for playing live 9x9 games and wanted to rank it up by playing some bot games to avoid unintentional sandbagging. I could not get a game with amybot-beginner (19k then) because that rank is more than 9 stones away from 6k.

"This is how it works unless it has been changed in the last 3 weeks.

“But this is old news.”

1 Like

I believe I was the first to note that starting players higher up had the ancillary advantage of shortening the life of alt sandbagging accounts. I did NOT offer that as an explanation or primary argument for starting players higher. Perhaps you heard the idea second-hand from someone who garbled what I said. The reason for the abandonment of letting players pick their rank, from what I heard from older moderators, was that it created way too much work for mods getting requests to adjust players’ ranks. If this isn’t true, then I was misinformed.

4 Likes

Why?
I don’t think It’s about fairness but about fun and what each player wants.
The system has to be made to answer different expectations so that dan players will not climb from 30k and beginners be crushed by 15k. Same time the actual system can still exist in between.

It’s the system to fit the players, not the reverse.

2 Likes

Did abandonment of picking initial rank happen before or after switching to Glicko?

The rank resetting workload for mods might be already fixed by Glicko’s faster settling of self-initialized provisional ranks.

But when both rating system changes were deployed simultaneously, this would not have been tried out in practice.

2 Likes

I’d solve that with:

  1. People can pick starting rank but it still has very low confidence so will adjust quickly
  2. Better help and instructions on the pick your rank page so that they get it right. You could even have a few tsumego.
  3. A policy that admins simply say “No, make a new account and choose better” to such requests and stick that in the FAQ.
3 Likes

Thank you for your reply Conrad :slightly_smiling_face: But I must protest that I’ve been quoted out of context. A more complete quote would be (emphasis added):

Thank you for noting that the 12 kyu is wrong :+1: If I understand correctly, it should be 12 kyu because of humble rank but that’s broken so it’s actually 6 kyu. So the situation is even worse than I thought! The important point is though that new accounts are starting at this high rank (12 kyu or 6 kyu, doesn’t matter) which is likely incorrect for most new accounts (too high for beginners, too low for dans). But there is no such restriction in Glicko-2 that says this must be the case, so why are we doing it? :face_with_raised_eyebrow:

I agree with 2 :+1: But actually 1 is unfair as beginners have to lose a bunch of times against much stronger players before they get a rank.

This would be crazy! :flushed: I’m not suggesting that. New accounts of IRL Dan players would declare an initial Dan rank and they would start from there. No climbing up ranks required. Only new accounts of IRL beginners would declare an initial rank of 30 kyu. A blank declared initial rank would revert to the existing default of 6 kyu.

Point 1 is exactly what I am suggesting :grin:

I agree with the “better instructions” part of point 2 :+1: But I think tsumego would be a step too far. Better to keep it simple and have something like “If you really don’t know, just leave it blank and your rank will automatically adjust after a few games”, in which case it would revert to the current behaviour and just start them from 6 kyu.

For point 3, no need to say “make a new account and choose better”. The answer is the current response of “just play a few more games and it will automatically sort itself out”. Declaration of an initial rank just sets a more appropriate starting value and so is just a way to skip the “lose 5-6 games against much stronger players” phase (or “play a bunch of boring easy games against weak SDKs” in the case of new accounts of IRL dan players). There is no need for any manual adjustment of rank by mods ever.

2 Likes

Here’s my opinion (I voiced something similar a few months ago, but it was dismissed as impractical for some reason).

I think there should be 2 options for new players -

  1. accept a 12kyu provision rank and play games until this rank adjusts
  2. take a quick / automated test and accept whatever provisional rank that test assigns.

Personally, I think that the test could be very simple - one could even take bits of the already-existing LEARN TO PLAY GO structure, and adapt those to tests. You could have simple things like

  • black to play to make two eyes (simple false eye tsumego)
  • look at a finished game and mark the dead groups (basic knowledge of life/death)
  • some very simple tsumego

My guess is - given the types of mistakes beginners routinely make - these tests would quickly catch 90% of the 25-30kyu players who arrive, assign them the appropriate TPK rank, and send them out to find players of similar strength.

For that remaining 10% that passed the first 3 questions would then receive 4 increasingly complex tsumego (15k, 10k, 5k, etc) until they made a mistake, and then that would be their provisional rank.

This entire process would only take new arrivals like 5-10 minutes tops, instead of the much greater time/effort required for someone’s rank to stabilize if they’re faced with playing multiple games (especially 19x19, and especially correspondence - it’s good to remember that a LOT of beginners are intimidated by live games and don’t want to be rushed…)

When I first found OGS in 2016, I mistakenly assigned myself a rank of 15kyu because I had finally beaten GnuGo on the lowest AI level, and I had been told that this was roughly equivalent to a 15-17kyu. I was WAY off. I was actually a lot more like 23-24kyu, and some very nice mod changed it to 25kyu for me. Having to play 12-15 kyu players was very confusing and intimidating, and it almost scared me back into my “only play against bots” hole rather than engaging with the site…

4 Likes

I would just let them integrate the current system.

It would be great enough to simply care of beginners as a bit of specific case who will never have the tools to even understand what’s happening to them.

Even if it’s still tough time, for these 10% they may have more philosophy and perspective to stand the rush and stay playing.

Another way would be to assign coach to full beginners. These coach don’t have to be 6d but be able to master bundaries, basic L&D, shichos… maybe 15k min? Coach will ensure that the beginner understand what is needed to not be a full beginner anymore, and then unlock the access to the rating system. Until then, playing between beginners would be promoted.

Coaching would surely solve a great part of the sandbagging problem too. And it can create more social links in the community. It could boost the image of OGS being the place to come to start playing go among all the go servers.
It should be confirmed to me but my guess there could be much more potential coach as newcomers. So it’s just a matter of organizing the system.

3 Likes