|
On July 20 2016 02:13 Excalibur_Z wrote: That's very interesting. Thanks! I really like following Bronze to Master journeys since it helps to get a good idea of the speed of progression.
It took an absurd amount of games. The score was like 140-11 and ~23 hours. Here's the VOD if you wanna skim over it.
|
You see this if the opponent is doing placements (presumably not their last placement):
I haven't figured out how random team games work yet. Does it take a straight average of the team?
|
the bonus pool is spent really weirdly now, anyone understands ?
|
United States12175 Posts
On July 20 2016 02:51 LordOfDabu wrote:You see this if the opponent is doing placements (presumably not their last placement): I haven't figured out how random team games work yet. Does it take a straight average of the team?
Yeah their final placement shows their new MMR and the MMR change for that game. I've been watching the VOD off and on today and for that -450 game, it was that player's second game of the season (which means it had to be unranked or else he would have seen a screenshot like the above). Really weird, because the math is probably right for a provisional state, but the fact that a 2499 was matched with a 3600 is crazy to me (that has to translate to 100% win probability).
For Random Teams, the matching uses your team's average against the enemy team's average (so 3000+3500 vs 3100+3400) but the rating change is applied individually against the other team's average (so for you it would be 3000 vs 3250 and for your teammate it would be 3500 vs 3250).
|
United States12175 Posts
On July 20 2016 02:32 pundurs wrote:Show nested quote +On July 20 2016 02:13 Excalibur_Z wrote: That's very interesting. Thanks! I really like following Bronze to Master journeys since it helps to get a good idea of the speed of progression. It took an absurd amount of games. The score was like 140-11 and ~23 hours. Here's the VOD if you wanna skim over it.
I took the liberty of going through the match history in its entirety here: https://docs.google.com/spreadsheets/d/1w-6sDS3POljVJNhXa_3Cq2kKE7_db-zQqMMSk1fPWAE/edit?usp=sharing
As we can see here, the closer the MMR values of both players, the closer the rating change for that game gets to 21-23. The outliers from games with unranked opponents also show an MMR gain of 46 (for 1101 and 835 rating difference games), which must be the maximum (the minimum being 0).
There's probably a lot more to be gleaned from this dataset, such as learning the relative trends of system confidence (you can see in the second chart that the rating difference gradually closes).
This is pretty fun and cool stuff. Thanks again for sharing!
|
Very interesting data to look at. Thanks for putting in the insane work laying it out!
|
It's quite weird. I got placed in gold with 4200 mmr playing diamonds and masters. If I won I gained around 50 points, when I lost 80-110.. lolololo
Just playing casul anyway hehe
|
The GM promotions don't make any sense.
You can see that at the time of promotion, there are people on Contender with MMR more than enough to be promoted, who meet the activity requirement (10 games in 3 weeks) that don't get promoted into GM, but people below them do.
|
United States12175 Posts
Oh damn, I forgot to watch the transition. I'll try to do that tomorrow and see.
|
On July 20 2016 16:19 Excalibur_Z wrote:Show nested quote +On July 20 2016 02:32 pundurs wrote:On July 20 2016 02:13 Excalibur_Z wrote: That's very interesting. Thanks! I really like following Bronze to Master journeys since it helps to get a good idea of the speed of progression. It took an absurd amount of games. The score was like 140-11 and ~23 hours. Here's the VOD if you wanna skim over it. I took the liberty of going through the match history in its entirety here: https://docs.google.com/spreadsheets/d/1w-6sDS3POljVJNhXa_3Cq2kKE7_db-zQqMMSk1fPWAE/edit?usp=sharingAs we can see here, the closer the MMR values of both players, the closer the rating change for that game gets to 21-23. The outliers from games with unranked opponents also show an MMR gain of 46 (for 1101 and 835 rating difference games), which must be the maximum (the minimum being 0). There's probably a lot more to be gleaned from this dataset, such as learning the relative trends of system confidence (you can see in the second chart that the rating difference gradually closes). This is pretty fun and cool stuff. Thanks again for sharing!
From what I can tell it looks like (approximately):
Change in MMR = +/- 22 + 0.025x(opponent's MMR less your MMR))
obviously '+' if you win and '-' if you lose.
The data doesn't fit perfectly.
So if you play someone with 100 MMR more than you and you win you should get 22 + 0.025*100 = 24.5 increase in your MMR (maybe rounded up to 25). Your opponent's MMR falls by -22+0.025x-100=-24.5.
Data fits much better the more games that are played. Where there are large outliers it's in the first half of the dataset mostly.
Interesting that it takes over 140 games to get to 5000 MMR which is probably below this person's true level. And it takes 80 games to get to 4000 MMR (my level) which is way below this person's level.
|
United States12175 Posts
MMR doesn't use rounding. All those decimal values are preserved internally (you can tell because the tiers cover equal MMR ranges but some tiers are off by one, so Gold being 3160, 3266, 3373, 3480 has a gap of 106, 107, 107 but are actually 3160-3266.66, 3266.66-3373.33, 3373.33-3480 for 106.66 each) even though the whole numbers shown on the client are truncated. So, decimal values for MMR gains are very possible and very important when deducing exact rating calculations.
The formula itself would have to be exponential. Something like 46 * (1 / (1 + (10^(-(rating_diff)/800)))) where going from rating_diff 100 to 200 has a bigger impact than going from 600 to 700.
The 140 games to 5k MMR (this account reached Grandmaster after the final game, if you were wondering) was only because it was a borrowed account that started from 2k MMR, below the Bronze floor. Placement matches swing your MMR pretty widely, probably 100-150 per match, which never puts you that far away from either end of the spectrum if you're starting a new account, since the seed MMR is around 3200-3300.
|
Yeah, it can't be completely linear. But alot of the data looks very linear.
|
In the AMA, they said it was like TrueSkill.
Under TrueSkill, your change in MMR depends on your MMR uncertainty as well as the difference between the 2 players MMRs.
And the relationship between the change in MMR and the difference between the 2 players MMRs is nonlinear, but approximately linear as long as the winner does not have significantly larger MMR than the loser.
|
|
United States12175 Posts
Yeah it's a result of that bug where the league boundaries shifted way way up for a few hours (Gold 3 going up to 5000 or so). In order to fix them, rather than rolling back to the old values, they just recalculated. It's going to make that page a little inaccurate since it means replacing filled cells, which isn't going to get quite as much crowdsourcing support as if those cells were empty at the start of the season.
|
What bug? How does something like that happen?
|
I think it's more fun to play. It'll encourage you more not to "grind" many games, but to play only 3 per day, until skill is dwindeling.
|
Hey Excalibur, do you have current ranges of the leagues based on population?
They said they were increasing the size of masters.. doubling it from 2 to 4% But myself and several others I know who were low masters are now tier 1 diamonds.. I'm personally a long way from masters where as prior to the ladder reset I was masters the previous 5 or 6 seasons
I'm not upset or anything.. just wondering if it's actually 4% now or if it's the 2% claim that was inaccurate
|
United States12175 Posts
On August 03 2016 01:13 paralleluniverse wrote: What bug? How does something like that happen?
It almost seemed like someone pushed the recalculation button but only for the top 1% or so of players. Master 1 changed to 6500-6611, Master 3 was 6280-6390, Diamond 2 was 5746-5973 (NA) 5760-5933 (EU). Then apparently it was erroneously pushed to the live servers. This happened on Friday afternoon and lasted for about 2-3 hours.
On August 03 2016 01:40 Ignorant prodigy wrote: Hey Excalibur, do you have current ranges of the leagues based on population?
They said they were increasing the size of masters.. doubling it from 2 to 4% But myself and several others I know who were low masters are now tier 1 diamonds.. I'm personally a long way from masters where as prior to the ladder reset I was masters the previous 5 or 6 seasons
I'm not upset or anything.. just wondering if it's actually 4% now or if it's the 2% claim that was inaccurate
The targets are 4/23/23/23/23/4. What that means is that the 96th percentile of users last season had an MMR of 4640 or higher, so that's the new boundary for Master 3.
One thing they mentioned in some Q&As is that they were messing with the percentiles for Master league during HotS/LotV without revealing their new target. They tried 6%, they tried 8%, but we continued to operate on the assumption that it was 2% the entire time because we didn't have any updated information and the community's existing tracking methods weren't aligned with what they use internally, so we didn't know how close we were. We don't know what the target was before the ladder revamp, but we know what it is now.
|
i'm playing 2v2s on WoL with a customer of mine as my team mate. and we get matched up with people in our tier and division all the time. it has created several rivalries with other teams. Kramerica Industries and Ball0 are in a war to end all wars.
i'm trying to get her to play HotS or LotV but she says its too complicated.
|
|
|
|