Grits N Gravy, on 13 November 2013 - 04:39 PM, said:
You can just convert the scores run a global K factor of 30-40 for a period then go to much a lower K factor. Or use the old system while you build up new data then roll out the new system once you are properly seeded.
I'd be a big proponent of this. I'd understood that they were using a 5 point base scale, not 50, which I considered perfectly reasonable. Slow, but reasonable. 50 points scaled up and down for relative value is too much swing. I know that I will hit 20+ matches some days and have over 5500 matches total. I should be nested almost immovably in my Elo target but instead, if it's 50 points per match, I could be swinging 1,000 points in a DAY. I've had 16 match losing streaks before. The difference between an 80 point and an 800 point swing is night and day. Over 200 matches 16 matches can easily get evened out but if I'm swinging 500-1000 points in Elo in a day.... well, the concurrent experience for the player is pretty wacky.
Grits N Gravy, on 13 November 2013 - 04:39 PM, said:
There is less total swing in Elo scores in a Gaussian implementation, coupled with a lower K factor. The points just dont trade hands as quickly. So if someone spends most of their time pugging and ends up 2000, it takes much more time for their rating to drop while they solo.
While they do solo, they will most likely be placed on a team with a 4 man, fighting another team with a 4 man. As a result of our narrower match maker. Which, results in more premade vs premade as they are more isolated from the solo drops. Even while soloing dropping our castaway 4 man dropper is more likely to see even matches.
This also reduces Elo drift. As players who usually quit do so with less Elo points than they started with.
The issue though is that a player who's earned a 2,000 Elo with his premade but pugs at an 1600 value is going to skew matchmaker results. If he's dropping with mostly premade teams what's going to happen is his team is going to be off by 400 points in value however you cut it. In fact this gets *worse* with a gaussian implementation - almost all the pugs on both sides will be people who earned their high Elo in premade fashion and are now pugging.
The problem here is that premades are treated as a single Elo amalgam made up of the players amalgam plus a modifier meant to represent the benefits of playing in a premade.
So suppose you've got 2 4mans on each team plus 4 pugs and the match estimated Elo score is 2000.
Team 1:
4man amalgam Elo - 1900
#1 actual score 1700
#2 actual score 1900
#3 actual score 1500
#4 actual score 1700
average 1700
premade bonus 200
amalgam Elo 1900
2nd 4man amalgam Elo - 2100
#5 actual score 1900
#6 actual score 1700
#7 actual score 2100
#8 actual score 1900
average 1900
premade bonus 200
amalgam Elo 2100
Pug # 9 Elo - 2200 - normally premades, without his team his value is 1800
Pug #10 Elo - 1800 - normally premades, without his team is value is 1600
Pug #11 Elo - 2000 - normally premades, without his team is value is 1900
Pug #12 Elo - 2000 - god tier pug who never premades *HIGHLY UNLIKELY*
Matchmaker Elo calculation:
Premade 1 1900 + Premade 2 2100 + Pugs 2200+1800+2000+2000 = 12,000/6 = 2,000.
Actual Elo value because of the pugs who are mis-calculated = 1883
Almost a 200 point swing.
Make sense? People perform better in teams than they do solo. Different skill set even, you know who you're with and how they'll act. You move in an organized group, you call targets on coms and you focus fire which multiplies your effectiveness. When pugging you have none of those benefits. In fact many of the builds that you use in a premade are only moderately effective when pugging. Ask any poptart or sniper or LRM boat. Without support your success plummets. Most rock solid 4mans will tell you that when they pug it's far, far more difficult.
The Gaussian model will skew higher Elo matches where you pretty much need to be dropping 4mans to get that level of success disproportionately because of this. Every pug in those matches is pretty much guaranteed to be rated higher than his skill represents. Even more bothersome is when these people get pulled to fill lower/higher Elo matches and a 'gimp' is pulled to offset the advantage they represent. Remember, not all premades are the same. There are plenty of premades that are just friends getting together without comms. Most fall in that description and they'll flesh out just about every band of Elo and require pugs to fill in the matches and inevitably to get weight matching and the like it'll have to reach into the +/-1 range to pull people. So it'll pull +1 pugs who normally premade who may be rated as much as 200 or more points above their actual skill while pugging, then it'll pull someone -1 who's legitimately 200 points below the Elo mean for the match and that team will end up 400 points out of variance because one guy is ranked incorrectly due to his difference in premade vs pug performance.
Are you getting what I mean? The difference for a lot of people between premade and pug performance can be
hundreds of points of Elo performance. The higher the Elo the more likely this is. When the matchmaker does need to pull further +/- scale to fill a match it's going to be pugs being pulled to fill a match in accurate weight classing so when it pulls one or two widely miscalculated players it's going to effectively double the miscalculation when it pulls someone from down scale to offset them and balance the team total.
It's an inherent problem that magnifies as Elo rises and has a literally exponential statistical effect on match results. Everyone in that match gets their results skewed because the Elo prediction is actually off by 10 or even 20%. While I agree that if it were isolated cases a lower k-value would help balance these out but most premades pug at least some of the time. That means that above a certain Elo band close to 100% of the pugs are going to be incorrectly ranked. Even if only 33% of any matches population is pugs that's going to throw Elo estimates off by almost the same % because the tighter Elo scale grouping that's such a boon for the middle 80% of players essentially tosses the top 10% to the wolves.
If pug Elo is separate and the highest Elo scores are essentially only possible in premade teams it's still going to involve having to pull pugs to fill matches in higher Elo games but at least Elos ability to accurately calculate values for each team will be preserved, ensuring that the resulting Elo impact on everyone is as fair as possible. The top tier matches are always going to have balance issues but at least this way you're giving the matchmaker the most accurate possible data from which to predict win/loss and award or remove points accordingly.