New AllegSkill ranks

[zombywoof]

QUOTE (cashto @ Jul 16 2009, 05:17 PM) A more serious problem is whether a (21) is worth the same as, more than, or less than three (7)s. Should a team of ten (21)s and ten (0)s be able to beat a team of twenty (7)s with 50% probability? Any schema to compress the upper range needs to consider this. IIRC, in the Glicko system (which could be considered a variant of Trueskill for 1:1 matches), a person of rating "n" is defined as someone who beats someone "n-400" 75% of the time.

That's the goal of allegskill, at least. Since it predicts the results of something like 90% of matches (or is it 80%?), I'd say it does a damn good job of it.

QUOTE So then, should a team full of (20)s be able to beat a team full of (17)s with the same probability as a team full of (10)s beating a team full of (7)s?[/quote]
No, but that's because twice of 7 is 14

[Broodwich]

look, people who suck are ranked higher than people who dont, and no amount of "better rank resolution" is going to change that. like allegskill, i think it did a better job of ranking people than helo did, and i respect baker for all hes done for the community. But it doesnt get everything right, and it never will. Pretending this is some "math land" problem and thinking that wider range of rank numbers will reflect people's skill more accurately is totally retarded. Plus add into that the fact that TE is so busy he cant even (in Baker's words) "run a 5 minute" script for the doubles tourney, and that almost every change ever made has run into errors, and needed an administrator to fix them, is really just asking for trouble. We dont need it

[Pedowich]

QUOTE (cashto @ Jul 16 2009, 08:17 PM) IIRC, in the Glicko system (which could be considered a variant of Trueskill for 1:1 matches), a person of rating "n" is defined as someone who beats someone "n-400" 75% of the time.

QUOTE (phoenix1 @ Jul 16 2009, 08:59 PM) No, but that's because twice of 7 is 14

He's saying its NOT scalar, as in an increase of 3 rank means a probability of 75% to beat someone 3 ranks below you. So a 10 v 7 would have same probability as a 12 v 9 or 17 v 14. `yy

[zombywoof]

QUOTE (Broodwich @ Jul 16 2009, 07:13 PM) look, people who suck are ranked higher than people who dont, and no amount of "better rank resolution" is going to change that. like allegskill, i think it did a better job of ranking people than helo did, and i respect baker for all hes done for the community. But it doesnt get everything right, and it never will. Pretending this is some "math land" problem and thinking that wider range of rank numbers will reflect people's skill more accurately is totally retarded. Plus add into that the fact that TE is so busy he cant even (in Baker's words) "run a 5 minute" script for the doubles tourney, and that almost every change ever made has run into errors, and needed an administrator to fix them, is really just asking for trouble. We dont need it

This is exactly what I've been saying since the beginning. AllegSkill is NOT "this person is better than the other." It's allegiance IQ, more or less. Richard Feynman (educate yourselves) was a genius, and had a stellar IQ of around 118 iirc. Didn't make him stupid. Likewise, Spawn, one of the best scout pilots you could have on your team, is ranked as an 8 (or is it nine?) He's a better pilot than I am (12). AllegSkill is "This person is PROBABLY better than the other." It's still up to using your brain in the end.

[cashto]

QUOTE (phoenix1 @ Jul 16 2009, 05:59 PM) That's the goal of allegskill, at least. Since it predicts the results of something like 90% of matches (or is it 80%?), I'd say it does a damn good job of it.

I have my doubts about that statistic, because if you think about it, it's more a measure of how stacked the average game is than it is a measure of how good AllegSkill is. Imagine if every game were horribly stacked. In that case, a good ranking system could tell you with near-100% certainty who was going to win. What if every game was moderately stacked? Well, then a perfect ranking system would be right most of the time, but every once in a while, the underdog team would pull an upset. If every game were perfectly and equally matched, well, the most perfect ranking system would only be right 50% of the time because every game would essentially be a coinflip.

I just find it hard to believe that 80% of Allegiance games are so horribly stacked that you can tell the winner 100% of the time. Or that 100% of Allegiance games are so stacked you can tell the winner 80% of the time. Or whatever the mix of the two is.

But back to the original point -- I think in most games, you can pair off most of the players as being of roughly equal rank, until you're left with a handful of players which determine the rating delta. The situation of 10 vets and 10 newbs vs 20 voobs happens very rarely if ever. And honestly I don't know how often Allegskill gets that right (then again, I don't have access to the game database).

[MrChaos]

No one has ever claimed that a 95% Confidence Interval means it's perfect. 5% of the time it will be wrong period dot. It also means your skill is across a range of values, with the higher the sigma the greater the range. So in order to increase the amount of times it gets everyone's skill level right (99%, 99.9%) you have to open the range of values it falls inside and it becomes a balancing act between your confidence interval range, the time to achieve results, and the accuracy to predict game outcomes, the REAL goal of the endeavour btw.

The claim that is being made is 80% of the time AllegSkills can predict the outcome of a game. Based on this it develops a listing of players per team to achieve the maximum balance possible using Mu and Sigma. Some days we play better and other days worse, this along with pure random chance means this is about as good as it gets since we aren't milling machines but human beings playing a game.

RANKS are for the great unwashed not AllegSkill and it's math. What would be best is for the community is to have the Allegskill version of autobalance on since it intelligently uses Mu and Sigma to balance games. BUT given the miniscule size of this community, we can not do this since there would be a wholesale revolt. I get it since there are people I simply won't play along side just like you. So we have to give the most accurate RANKS so commanders have as much info as possible to pick fairer teams. They aren't wrong just the best try at getting five lbs of info in a single number so the players have something to measure against.

I'd LOVE to have an intelligent discussion about things since there a number of grey areas on the implementation end IMHO (not they are wrong but there are some knobs you can tinker that may be beneficial in other areas) but mother $#@!er we can't get past giving the ranks $#@!ing names The change isn't arbitary but rather essential since it's artificially supressing rank CHANGES.

Go buy "The Dunkard's Walk. How Randomness Rules Our Lives" by Leonard Mlodinow who does a superb job of trying to explain why intuition, and mental gymnastics don't work when dealing with randomness and probability.


edit: While I am a little annoyed with Cashto, just a little, I did not write this response to him rather it's a generic one given we have this kind of discussion over and over and over again. It makes one defensive and poopie. It kind of like quesitoning Thalgor about the donations a while ago. It's insulting to him and come to find out he WAS a good guy. Hmmmmm wonder why Baker and TE keep putting off these changes

[ImmortalZ]

*removed big ass rant directed at the village idiots who don't know @#(!*

[fuzzylunkin1]

Cashto the less the numbers are changed the more accurate they are (even if it's a fractional amount).

[Correct]

QUOTE (cashto @ Jul 16 2009, 08:31 PM) I just find it hard to believe that 80% of Allegiance games are so horribly stacked that you can tell the winner 100% of the time. Or that 100% of Allegiance games are so stacked you can tell the winner 80% of the time. Or whatever the mix of the two is.

I could probably tell you the outcome of at least 70% of games based on team, settings and map. Most vets could [not rank vet, actual vets]

[MrChaos]

QUOTE (Correct @ Jul 17 2009, 12:12 AM) I could probably tell you the outcome of at least 70% of games based on team, settings and map. Most vets could [not rank vet, actual vets]

You actually can't but you think you can, read the book Eyeball. I mean no offense... you know whenever we are on you have to predict the winner at launch every single game your on no cherry picking, I'll keep track and after say 300 games we'll post it, maybe a 100. It will be an interesting experiment


If it's horriblely stacked it is 99.999999999999999999999999999999999% accurate, it's 80% accurate across the entire database home slice