EPT[G] Comprehensive SC2 League and Ladder Guide - Page 62
| Forum Index > SC2 General |
|
korona
1098 Posts
| ||
Excalibur_Z
United States12235 Posts
On June 18 2012 20:39 korona wrote: Excalibur_Z: As I understand the exact amount of games that Blizzard includes when calculating the moving average of MMR is not publicly known? Is there some rough guess how many games are included? This info becomes more interesting as through skeldark's MMR tool & NotThat's analysis we get better understanding how our relative MMR behaves. The tool also offers a moving average calculation, but it is only based on 5 games. That's right, it's not known. The tricky part about figuring out the moving average interval is that it's something that's completely hidden from view. The system uses a hidden slope threshold based on a hidden value with a hidden interval. In addition, there's a special rule which guarantees an instant promotion once your MMR crosses into 2 leagues higher than your current league, making tracking this a little more difficult (because you don't want to trigger the special rule and invalidate the test). For best testing results we would need an account which is able to get into a higher league but use the stabilization requirement to get promoted. | ||
|
VediVeci
United States82 Posts
On June 19 2012 00:26 Excalibur_Z wrote: That's right, it's not known. The tricky part about figuring out the moving average interval is that it's something that's completely hidden from view. The system uses a hidden slope threshold based on a hidden value with a hidden interval. In addition, there's a special rule which guarantees an instant promotion once your MMR crosses into 2 leagues higher than your current league, making tracking this a little more difficult (because you don't want to trigger the special rule and invalidate the test). For best testing results we would need an account which is able to get into a higher league but use the stabilization requirement to get promoted. I don't know how familiar you guys are with other ranking/matchmaking systems like ELO/Microsoft's Trueskill/Glicko, but the way those work (and I assume the way SC2 mmr works) is that you have your rating and it is updated after each game or at the end of the day based solely on the games you have played since it was last calculated. So in a way, every game you have ever played would be taken in to account, but in another sense, none of your previous games matter, just your current rating (and possibly another piece or two of related data). There are several reasons why this sort of system, based on Bayesian inferences and distribution functions, is advantageous. I can discuss these reasons, or more of the math/methodology behind the other systems if you want, but my educated guess, based on experience with these systems, is that this is the way Blizzard's system works. | ||
|
Excalibur_Z
United States12235 Posts
On June 19 2012 12:50 VediVeci wrote: I don't know how familiar you guys are with other ranking/matchmaking systems like ELO/Microsoft's Trueskill/Glicko, but the way those work (and I assume the way SC2 mmr works) is that you have your rating and it is updated after each game or at the end of the day based solely on the games you have played since it was last calculated. So in a way, every game you have ever played would be taken in to account, but in another sense, none of your previous games matter, just your current rating (and possibly another piece or two of related data). There are several reasons why this sort of system, based on Bayesian inferences and distribution functions, is advantageous. I can discuss these reasons, or more of the math/methodology behind the other systems if you want, but my educated guess, based on experience with these systems, is that this is the way Blizzard's system works. No, you're right, that is how Blizzard's system works too. However, other systems don't have promotions and demotions like SC2 does. What Blizzard uses to determine whether you should move to a new league is a moving average of your past games. So, while your uncertainty and your current rating and everything else is not influenced by your past games, your promotion opportunity is. They use this as a stability measure so that players who hang around the border of two leagues don't get promoted then demoted every couple of games, because they want you to remain in a division long enough to climb in rank and for that division to mean something. | ||
|
VediVeci
United States82 Posts
On June 19 2012 13:21 Excalibur_Z wrote: No, you're right, that is how Blizzard's system works too. However, other systems don't have promotions and demotions like SC2 does. What Blizzard uses to determine whether you should move to a new league is a moving average of your past games. So, while your uncertainty and your current rating and everything else is not influenced by your past games, your promotion opportunity is. They use this as a stability measure so that players who hang around the border of two leagues don't get promoted then demoted every couple of games, because they want you to remain in a division long enough to climb in rank and for that division to mean something. Oh that makes sense, apologies for the confusion. So it's known that they use only the past x games to determine promotion/demotion, or is that conjecture? I looked at the original post and didn't see a definitive source confirming that theory, but I might ave missed it. | ||
|
Excalibur_Z
United States12235 Posts
You need to be in the top 200 players according to a rolling average of your skill rating. That "rolling average" requirement applies to all leagues. We first learned about this at Blizzcon 2010, but were unsure of what type of moving average was used. There was a UCI lecture that the designer gave in March 2012 in which he mentioned that it was in fact an exponential moving average. | ||
|
Lysenko
Iceland2128 Posts
On June 19 2012 15:47 Excalibur_Z wrote: That "rolling average" requirement applies to all leagues. We first learned about this at Blizzcon 2010, but were unsure of what type of moving average was used. There was a UCI lecture that the designer gave in March 2012 in which he mentioned that it was in fact an exponential moving average. Is this lecture available online anywhere, as video or a transcript? | ||
|
Excalibur_Z
United States12235 Posts
| ||
|
VediVeci
United States82 Posts
On June 20 2012 00:37 Excalibur_Z wrote: Yes. It's fairly long but pretty interesting. http://www.ics.uci.edu/~develop/Lectures/CGVW031412_MP4 360p (16x9).mp4 Thanks to Not_That for finding this. Haha that was incredibly serendipitous, my sent me his notes to that speech an hour before you posted that and said it had been really informative. Thanks for the link! | ||
|
Tao367
United Kingdom324 Posts
| ||
|
Lysenko
Iceland2128 Posts
On July 03 2012 23:48 Tao367 wrote: Sorry, couldn't really find an answer anywhere, but does the system take into account matchups in anyway? I.E My pvt is good so I will face "better" terrans, whereas my pvz is weak so I'll face weaker zerg opponents? No, it does not. If you're notably weaker or stronger in one matchup, it just means your MMR will fluctuate more depending on against whom you wind up playing. | ||
|
skeldark
Germany2223 Posts
If we find the time we write a summary about what we found out. Also not_that wants to calculate the offsets again when we reach 100K games ( 80K at the moment) I thought again about the "promotion average" and im pretty sure we can not solve this one. I think it depends on the gauss variation of the user and this one we can not back-calculate ( or not with reasonable effort). But it can be that its just an past moving average. For masters it look like 7 -15 games above the promotion-line. Oh and i publish the data if i have 100k games after i removed all personal information from it. Sadly there are not many people who use it to make some interesting graphs. When i have more users i could perhaps even calculate the skill diffrence between the different server. | ||
|
Lysenko
Iceland2128 Posts
On June 20 2012 00:37 Excalibur_Z wrote: Yes. It's fairly long but pretty interesting. http://www.ics.uci.edu/~develop/Lectures/CGVW031412_MP4 360p (16x9).mp4 Thanks to Not_That for finding this. Finally got around to watching this. He did give a hand-wavy (but somewhat illuminating) explanation of how they got to "skill-adjusted race balance" information, which has been a frequent question when it's come up from David Kim's various talks on game balance. (For anyone looking for this, it's near the end, in the Q&A.) | ||
|
thurst0n
United States611 Posts
My question is more like "does the system match me against people that it thinks will give it a better idea of my skill" or in other words "Is the matching system intelligently matching me, or is it just random" I'm just thinking about win streaks (or losing streaks) If I beat someone who the system thinks I'm supposed to beat, that makes my certainty factor go up, AND my mmr goes up slightly(right?) So If I keep doing that, I know that if it's random eventually I'll play someone the system thinks is better than me, but will it try to match me with someone better sooner, assuming that I'm on a win streak, or someone worse sooner if it's a losing streak. My question is kind of all over the place, but hopefully someone understand what I'm trying to ask. If it's 100% random, then that answers my question as well. | ||
|
Lysenko
Iceland2128 Posts
On July 06 2012 22:24 thurst0n wrote: My question is more like "does the system match me against people that it thinks will give it a better idea of my skill" or in other words "Is the matching system intelligently matching me, or is it just random" I would recommend watching the video from the matchmaking designer's talk at UCI that Excalibur_Z posted. He discusses this at some length. Basically, it tries to find another player in the queue relatively close to you in skill. However, the matchmaking server tries to globally optimize the matches for everyone in the queue, which might cause you to be paired up with someone who's not the best match for your individual skill level. He also said (correct me if I don't get this right, I'm speaking from memory) that the most efficient way to find everyone's skill as fast as possible would be to match all the people whose skill levels were poorly known against each other, but since that would tend to yield a lot more uneven matches early in a person's play history, which would be un-fun for new players, they leaned toward matching people with uncertain ratings against people with more certain ratings. So, I think the answer to your question is: it's not random, though it is dependent on who's queueing at the same time as you. It's explicitly not seeking the match that will lead to the most stable rankings fastest, because that strategy gives more uneven (i.e. un-fun) matches to new players. One other point he made is that the hidden MMR changes pretty fast as a result of win or loss streaks, and (I'm inferring this) this may be why it's common to have a win streak followed by a loss streak, repeatedly, as one oscillates around one's "true" score. | ||
|
oGoZenob
France1503 Posts
So my question is, how much game does it take to get your mmr back on track ? I'm curently 50-2, and still matched against low-mid-top bronze. I was fun first, but it's kinda backfiring at me ^^ | ||
|
Lysenko
Iceland2128 Posts
On July 14 2012 06:49 oGoZenob wrote: ok little question here. I'm a diamond terran and i recently bombed my mmr down to bronze, to get motivated to play more (yeah, i know) So my question is, how much game does it take to get your mmr back on track ? I'm curently 50-2, and still matched against low-mid-top bronze. I was fun first, but it's kinda backfiring at me ^^ Recently, someone posted about a person who had done something similar and took about 130 wins in a row to get back to Master league. Your mileage may vary. | ||
|
oGoZenob
France1503 Posts
On July 14 2012 07:11 Lysenko wrote: Recently, someone posted about a person who had done something similar and took about 130 wins in a row to get back to Master league. Your mileage may vary. wow but the 130 win were to get back where he come from, right ? I just want to get matched against better player, as i'm still matched against bronzies after basically 50 straight wins | ||
|
Lysenko
Iceland2128 Posts
On July 14 2012 09:04 oGoZenob wrote: wow but the 130 win were to get back where he come from, right ? I just want to get matched against better player, as i'm still matched against bronzies after basically 50 straight wins I think that was 130 wins to get promoted to master league again, so getting to diamond may be short of that by a bunch. Still, to get demoted to bronze you had to fall pretty low in the league, so I'm not surprised that it would take a lot of games to make your way back up. Never fear though, all those wins ARE increasing your MMR. | ||
|
_mylittlepwny
United States9 Posts
On July 14 2012 09:04 oGoZenob wrote: wow but the 130 win were to get back where he come from, right ? I just want to get matched against better player, as i'm still matched against bronzies after basically 50 straight wins If you don't stop winning, I believe your MMR should ramp up very quickly since your uncertainty will be huge. You can probably get back to diamond in a similar buffer.. | ||
| ||
