MMR deviations can easily be distorted by a number of games against a race which you fare better against than others etc.
SC2 Ladder Analysis - Hidden Ratings. - Page 5
Forum Index > Closed |
balls84
United Kingdom53 Posts
MMR deviations can easily be distorted by a number of games against a race which you fare better against than others etc. | ||
PimpMobeel
120 Posts
[B]On November 01 2010 08:36 darmousseh wrote: . 1: The only people that play are those that are serious. 2: The skill gap between a grandmaster and an amateur is pretty small. Hmmm. | ||
BurningSera
Ireland19621 Posts
| ||
MasterJack
Canada215 Posts
On November 01 2010 22:45 balls84 wrote: Serious trolling going on here, but another reason chess ratings are so close is because every match is a mirror match. MMR deviations can easily be distorted by a number of games against a race which you fare better against than others etc. I disagree with this. You have to see your MMR as an "average" against all races. You compare to this everyone's average against all 3 races. This would only be distorted if you could pick your MU. Theoretically, with a perfect balance, there are no MMR race differences. It CAN be easier to achieve a higher MMR with 1 race (ie: 4 gate, or 3 rax all-in), but this doesn't discredit MMR. | ||
TERRANLOL
United States626 Posts
http://sc2ranks.com/team/3835609 Notice how the points go up consistently and the world rank jumps all over the place(just not straight up)? That's inflation | ||
Leviance
Germany4079 Posts
On November 01 2010 22:30 Merano wrote: @Leviance: You are not #125 player in europe. You are #125 of all teamliquid users in europe typing their name into that web form. Several 1000 other players much better than you are still unknown to the system. Ah I see that makes sense. Sorry, somehow I thought the page has all the b.net data ready like sc2ranks.com. | ||
emythrel
United Kingdom2599 Posts
the OP was actually quite accurate with those statements. The difference between an amateur and a GM in chess is more often than not simply experience. You can master the knowledge of the game quite quickly and there comes a point where simply being more experienced than your opponent will net you alot of wins. The difference between and GM and everyone else is the GM's can work out all the variations in their head, not that the GMs are actually majorly more skilled at the game. It doesn't take skill to work out variations, it takes experience or just having the right kind of mind for it (like the child prodigies that pop up every few years). Josh waitzkin (think i spelt it right) was 2200 rated by the time he was 15 but often lost to older, more experienced, lower ranked opponents. Chess is however not the best analogy for SC as there is only one match-up to master. And a limited amount of ways to move your pieces. SC2 has 6 different possible match-ups, an infinite numbers of ways to move your units and literally unlimited ways to group those units into armies. You have multiple ways to cripple your opponent, and many different ways to win... in chess you have only 2... mate the king or force your opponent into such a bad position they tap out. On topic, great website OP..... even tho its unlikely you have managed to hit the magic formula (i know for a fact you haven't because they showed it at blizzcon and it was waaaaaaay more complicated than your formula) you have managed to devise a way for us players to see roughly how we stack up, whether we are near promotion etc... and that should be applauded. Its a great first step towards having a place where we can go and find our true rating. edit: according to your OP i am very near promotion to Plat and have been for a while... guess i need to play a little more regularly. On a related note.... sc2ranks.com puts me just outside the top 200 gold players in europe (atleast on points) and almost top 600 in the world.... I know points mean fuck all, but its still nice to know of people in my league.... i am up among the best lol | ||
vanick
United States53 Posts
On November 01 2010 22:29 Catyoul wrote: What you have found though is how the different leagues compare, and how we can convert from one to another. That in itself is interesting enough. He actually did no such thing. He observed that players who have used their bonus pool generally have points between 1000-1500, except in Diamond where there are higher point values (because anyone good enough to rise above a certain number of points in lower leagues would get promoted). I'm not sure if his 1000-1500 observation is exactly correct (it looks a lot more like 1000-1600 in my brief look) but even worse, his 500 point offset comes from basically nowhere and is not actually justified as a solid assumption. As you said before, the formula put forth is just readjusting a player's earned points by a league offset. And given what we know about the volatility of MMR when a player is streaking, and that divisions have other adjustments made to the points there is no way this is even close to accurate. That is, using this formula would be more misleading than illuminating. This combined with his solicitation to Shadowed for ad revenue, and the topic title piggybacking on the work ExcaliburZ put into his thread, makes the whole thing stink. | ||
rkiga
United States44 Posts
On November 01 2010 13:15 Excalibur_Z wrote: The formula posted by the OP is also incorrect because it considers points even across all divisions, which we know from Blizzard not to be the case. I tried searching here and google for a citation but couldn't find one. Can you or somebody point me to a blizzard post confirming this? There was no citation in your two threads here on TL, and there was nothing in the Blizzard "Leagues and Ladders FAQ". | ||
Cheerio
Ukraine3178 Posts
division [bronze = 0, silver = 1, gold = 2, platinum = 3, diamond = 4] (division - 1) * 500 + (points) + (bonus pool remaining) + (1500 - total bonus pool) what a waste of time. Basicly for diamond you assume that MMR = 3000 + visible_rating - bonus_pool_used with pretty much no evidence. And what is even more funny is this one: First off, I need to clarify that this is only possible if it is true that your points will converge to your actual rating over time. Do you understand that by this remark you mean that for diamond MMR = 1678 + visible_rating (or otherwise your whole formula just doesn work and cant be applied) because over time the bonus pool will be all depleted? Tell me please what is the use of knowing your MMR if it ranks you in 90% cases JUST the same as your actual ranking does? For example i am a 358 platinum player with 1004 bonus points remaining. Currently all players have received 1322 bonus pool so my MMR is (3 - 1 ) * 500 + 358 + 1004 + 178 = 2540. For me at 2540 it means i am very close to getting promoted to diamond, however, probably won't be promoted unless i maintain that rating for a certain number of games or period of time. This is just lol. You are making conclusions while breaking you own assumptions. ANY newly promoted platinum will have nearly the same MMR like you do because all of you will have pretty much around 300 actual rating and nearly the same bonus pool. So one more conslusion from your thread is that any newly promoted platinum = diamond. Isnt that nice to know? Thanks for wasting my time. | ||
Excalibur_Z
United States12204 Posts
On November 02 2010 00:13 rkiga wrote: I tried searching here and google for a citation but couldn't find one. Can you or somebody point me to a blizzard post confirming this? There was no citation in your two threads here on TL, and there was nothing in the Blizzard "Leagues and Ladders FAQ". They've danced around the topic in the past, but if you look at the video of the Q&A from the Blizzcon SC2 Multiplayer Panel, that's where they really answer it, since they mention the "skill of your division". Vanick and I talked to the Senior Designer (the rightmost panel member) after the panel concluded and learned that points can't directly be translated across divisions, and that Blizzard wants the emphasis to be on your division rather than global to that end. He also said that when he "does the Top 200 every week, that's all by points, but with all the division stuff taken out" which makes it pretty clear that there's some kind of weighting or adjustment involved. | ||
Mendelfist
Sweden356 Posts
On November 02 2010 00:35 Excalibur_Z wrote: He also said that when he "does the Top 200 every week, that's all by points, but with all the division stuff taken out" which makes it pretty clear that there's some kind of weighting or adjustment involved. So what could be the reason for adding any sort of division weighting to your points in the first place? I can't think of any. | ||
NukeTheBunnys
United States1004 Posts
TL raped your bandwith Its interesting, and I think you have part of the system down, but I think the delta value comes into play to play when blizz does its top 200, still better then anything else we have | ||
Excalibur_Z
United States12204 Posts
On November 02 2010 00:46 Mendelfist wrote: So what could be the reason for adding any sort of division weighting to your points in the first place? I can't think of any. I'd speculate that if there are indeed "tiers" of divisions such that everyone in your division is roughly your skill level, then it makes a little more sense. So let's say that a new Gold division is created and its target MMR range is the top 20% of Gold (this is still speculation). That means that all 100 players within that division will, at the time of their promotion/demotion/placement, be in the top 20% of Gold. From there, you can get a fairly accurate ranking of who is better than who based on points in that division. That may not translate to the points generated by a 20-40% Gold division or a 40-60% Gold division, even though the points may be similar. Specifically, the designer said that at Team Liquid, we are the top of the top Diamond who focuses only on the top of the top Diamond, but that for 99% of the player base, the division system works properly. He said that the Master League is designed to address the issues that we're noticing now. The problems that we're seeing are centered around (1) Diamond being the highest league and (2) the skill gap between high Diamond and low Diamond being too great, which is sort of where the division system breaks down. Diamond comprises the top ~8% now of players, that's over 40,000 players on the NA server. Because of the way the skill curve works, the top 5,000 or so players may be vastly more skilled than the next-highest 5,000 in Diamond, and that the top 500 players may be vastly more skilled than the next 500. However, it's possible that some of those players could be in the same division, and that causes enormous rifts between the top and bottom of each Diamond division even though the skill ranges are supposed to be tight. | ||
Mendelfist
Sweden356 Posts
On November 02 2010 02:09 Excalibur_Z wrote: I'd speculate that if there are indeed "tiers" of divisions such that everyone in your division is roughly your skill level, then it makes a little more sense. So let's say that a new Gold division is created and its target MMR range is the top 20% of Gold (this is still speculation). That means that all 100 players within that division will, at the time of their promotion/demotion/placement, be in the top 20% of Gold. From there, you can get a fairly accurate ranking of who is better than who based on points in that division. That may not translate to the points generated by a 20-40% Gold division or a 40-60% Gold division, even though the points may be similar. I can see the point of making divisions of people with equal skill level, but not adding any division weighting, other than hiding the fact that the divisions are not equal (more obfuscation). Is that what you mean? | ||
zZygote
Canada898 Posts
| ||
Excalibur_Z
United States12204 Posts
On November 02 2010 02:37 Mendelfist wrote: I can see the point of making divisions of people with equal skill level, but not adding any division weighting, other than hiding the fact that the divisions are not equal (more obfuscation). Is that what you mean? Obfuscation could be part of it, he declined to state the reason. He just said that "we didn't actively seek to screw over sites [like SC2Ranks], the division system just happens to do that". How the divisions actually work in that context is anyone's guess. | ||
darmousseh
United States3437 Posts
The only reason i can say anything at all and the only reason i posted is because blizzard confirmed that "your points will converge towards your rating over time". And that playing more or less games will not affect your rating at all (but possibly your displayed points). The way i made this discovery was by noticing that the number of people in silver - platinum below 1000 or above 1500 was really small (i still haven't completely decided if it is 1500 or 1600). Given what has been said about the MMR in the past, that they look at the moving average and check if it is above a certain threshold, and then promote you means that The top of one division must be very close to the bottom of the division just above it. It is highly unlikely that blizzard is doing anything too fancy and the more i think about it, the more it seems like it's the best way to hide your rating information. A lot of people seem confused as well, MMR has two components mean and sigma. After you have played roughly 50 games (well, it really depends how they set it up), then your sigma will be small enough for your rating to have meaning. If you are 1500 with 150 sigma (which is a typical starting point for chess), then you are somewhere between 1200-1800. That doesn't tell you anything really. But if you are 1500 with a 25 sigma, then you are somewhere between 1450 and 1550 which is a lot more information. Other important thing is that some people are misunderstanding how i came to these conclusions. There are 2 ways of proving something like this. 1. Deduction 2. Hypothesis and Verification. 1. We can only deduce so much. The information blizzard has given us leads me to believe they are using trueskill (or some other bayesian formula), that points do reflect your rating (to a certain degree), that you will eventually earn all your bonus points (so we can just negate them), and that the top of one division are very close to the bottom of the next division. This by itself was not enough to come up with anything, however, the thing i noticed was the phenomena of the 1000-1500 (or 1000-1600...still can't decide) in ALL divisions for players with at least 200 games(so entire bonus pool used). So i thought, what if blizzard is being retarded and simply offsetting the divisions by some points? So i decided to go with 500 for testing purposes. Division offset * 500 + points. I took some random data from all of the divisions, applied the formula and checked the distribution and guess what, it was perfect (for those with at least 200 games). At this point I am very confident, if you have used your entire bonus pool it is easy to compare yourself to other divisions. Then i assumed that what if i just added the bonus pool remaining for those who haven't completed the games. So i added the bonus pool (estimation since i don't actually have the data. the estimation was 12 * wins used) and added them to the grouping and the distribution was still equal except for people under 50 games. So, blizzard being blizzard I also assumed (and probably correctly) that displayed points probably reflects either mean, mean + 2* sigma or mean - 2*sigma (for the 95% confidence). I tried all 3, but noticed that the mean + 2 * sigma produced more even results. It was only after that that i decided that this might be correct. I did some comparisons between myself and a few people and saw that the expected points was correct for our rating differential, etc. Things to understand about my formula. This does not give you your actual MMR. It gives you an approximation of your MMR so that you can compare yourself to someone else. There is also another possibility that i have considered and if true, i would need to adjust accordingly. It is entirely possible that the groupings are scaled. What i mean is this. I said silver is 1000 silver = 1000 rating 1500 silver = 1500 rating, but it is possible that 1000 silver = 1000 rating 1500 silver = 1400 rating. So points in silver would have to be adjusted for that scale. I considered this a real possibility when i was analyzing this, but came to a different conclusion for a different reason. Blizzard mentioned that the groupings didn't turn out exactly how they wanted. Bronze was too big and diamond too small. If this is true it means that the groupings that blizzard used were incorrect and do not self adjust during the season, therefore they are most likely static based on the ratings that came out during the beta as was mentioned in a recent post that i linked in the OP (The one referring to the 0 bronze favored problem) and that post basically confirmed what I was thinking. In the end the equation is very simplistic and can be boiled down to division offset + points for those who have used the entire pool, but i added in the bonus pool and modifier for those under 200 games. There is one last reason i have to believe that this is correct. The numbers that i end up with are WAAAY too similar to chess and i don't think its a coincidence. xbox live also uses a bayesian algorithm but their scale is 1 to 50 (and just round the decimal point off for display). Everything from the 1500 for the group maxima to the 150 point sigma that worked (i tried everything from 0 to 250, but 150 gave the best results) to the 50 games (also tried from 0 to 100). Side note I spoke to shadowed from sc2ranks that i was posting a possible formula for MMR and that I wanted to make sure that i could use his api and create a website to post the information. If anyone needs to discuss this further please send me a PM, don't troll the thread. | ||
darmousseh
United States3437 Posts
On November 02 2010 02:09 Excalibur_Z wrote: I'd speculate that if there are indeed "tiers" of divisions such that everyone in your division is roughly your skill level, then it makes a little more sense. So let's say that a new Gold division is created and its target MMR range is the top 20% of Gold (this is still speculation). That means that all 100 players within that division will, at the time of their promotion/demotion/placement, be in the top 20% of Gold. From there, you can get a fairly accurate ranking of who is better than who based on points in that division. That may not translate to the points generated by a 20-40% Gold division or a 40-60% Gold division, even though the points may be similar. Specifically, the designer said that at Team Liquid, we are the top of the top Diamond who focuses only on the top of the top Diamond, but that for 99% of the player base, the division system works properly. He said that the Master League is designed to address the issues that we're noticing now. The problems that we're seeing are centered around (1) Diamond being the highest league and (2) the skill gap between high Diamond and low Diamond being too great, which is sort of where the division system breaks down. Diamond comprises the top ~8% now of players, that's over 40,000 players on the NA server. Because of the way the skill curve works, the top 5,000 or so players may be vastly more skilled than the next-highest 5,000 in Diamond, and that the top 500 players may be vastly more skilled than the next 500. However, it's possible that some of those players could be in the same division, and that causes enormous rifts between the top and bottom of each Diamond division even though the skill ranges are supposed to be tight. I'm almost 100% certain that someone asked this during the beta and they confirmed that once you get into a division you are stuck there. Weighing divisions would be useless. If i'm correct then master league will start at 500 points higher than diamond assuming they don't do any scaling in the future (read my most recent reply). | ||
darmousseh
United States3437 Posts
| ||
| ||