Alternative balance

.csv

1847 team games played in last 60 days. No bots, no FFA, no missions, game must have at least 4 but no more than 24 players, and none of them can have EloWeight < 5.

Format is to Kyubey specifications, except "team 1" (not actually numbered in the file) is specifically defined to be the winner.

Note: the Elo values are as of time of writing, not from when the game was actually played.

EDIT: Damn, this thing is clearly defective; there are 117 FFAs in the list. I don't know how they got in, but you can just delete all the entries with winning team Elo variance == 0.

well even from basic data analysis we can see that variance is a verry good indicator of what team will win the game

banana_Ai http://en.wikipedia.org/wiki/Variance i think
but now looking at the numbers, im not entirely sure
KR i think your getter is bugged

well 10 11v11 games to work with is good enough for a test batch
(after pruning off all the weird 12v11, 12v10, chick/bots, thats what i was left with)

im going to sleep soon, i dont feel up to making something that will get the variance of all the games ever played no mater how many players (excluding 1v1 to 3v3 games) right at this moment, that is for tomorow

nevermind earlier, that is variance ie. r^2 i blame it on being 5 in the morning!
do sqrt(variance) for avrage distance from mean

@KingRaptor

if variance between teams is a more significant determining factor for win chance then avarage ELO curent system is completly usless (since it does not take into acount variance)

if not thie thread will be flooded with gifs till it crashes and burns

all future threads may be pointed to this thread and subsequently flooded with gifs

yea, stdev is totaly irrelevant

if anything, considering that the winning team had on avrage 27 points more ELO stdev then loosing team having a higher stdev is counterintuitavely a beter indicator of who will win then elo diffrance

afres some copious deleting of games that were 2v2 or less i got down to 1415 games

winers had:
avg elo of 1651, and stdev of 202

loosers had

avg elo of 1628, and stdev of 178

the diferance between loosers and winners was on avrage:
22 ELO with, and 24 stdev (this sugests that larger stdev is better for winning)

intrestingly enough, of the 525 teams that had less elo then enemy and still won:
1625 avg elo
37 points less on avarage then enemy elo
which was 1662 on avarage

had 201 stdev for the winners, and 186 stdev for the loosers


the other games (where winning team had more elo)
had 204 stdev for the winners, and 173 stdev for the loosers


no notable corelation between stdev(win-lose) and ELO(win-lose) to be found

make of that what you will, im going to sleep

Squigglier than a climatology graph.

Alright nabs, you got your fucking stats. Now go prove the balance is inefficient.

There is a very simple way to do it. All you need is to write a predictive script that can predict the probability of one team winning. You then run it through the data, making bets on some team.
If it can gain significant winnings, then the script is able to detect unfair games where Springie couldnt, thus proving Springie's inefficiency.


Now to formalise the betting part:

For each game, you can bet from 0 to 100 points on a team. A lost bet pays out 0. A won bet pays out the bet value plus 2*(winProb), where winProb is the winning team's probability to win according to Springie.

I am pretty sure you could easily prove that games with odd playercounts are biased against smaller team and 2coms dont help enough, and that the elo difference is more sensitive than springie's average predicts (as in, 10 avg elo difference results in bigger win chance than 51.4).


But these are fairly obvious. I want you nabs to come up with something more complex to prove that springie balance is fundamentally wrong, as you keep claiming.

I think the above summarises the problem perfectly. Sfireman is an effect, not a cause -- ZK becomes very unrewarding at high elo levels.

quote: balance is drunk like sfireman, his elo is near 1900 I think, but it worth -1900.anyway the balance system is very drunk , 1 pro with 2 newbie lose to 3 average player

This is the beef. So long as this happens, teamgames will forever become exponentially less fun as your elo increases.

I think it is important to fix this, as it drives good players away. Certainly when I was 2000+ elo I felt nothing but frustration from teamgames of any real size.

quote: I think it has been proven that despite the official claims, 1v1 elo still affects team elo

Team elo hasn't been changing from 1v1 games for at least a week, maybe two now. It hasn't been recalculated though, so it still is heavily influenced by past 1v1s.

quote: Now to formalise the betting part:

Note that the proponents of a balancing change do not dispute fair victory chances.

That entire argument is that high stdev causes unpleasant while still winnable games, which implies that stdev should also be optimized against. Current algorithm completely disregards the second parameter, so we can tentatively consider it to sacrifice stdev in favor of win-chance.

Thus the change-proponent side should produce an algorithm that optimizes both win chance and stdev, which will likely lead to reduced accuracy of win-chance optimization. The question is how much winchance-fairness will be lost to procure extra bits of stdev-fairness.

Unfortunately, you cannot resolve the efficacy of such an algorithm simply by betting on historical data, because you cannot rearrange the teams and see what happens. At best, you can calculate "what i would have done" for each of those games and see if they are indeed fairer (either/or by winchance or stdev)

Also note that any bets you make on this dataset will be tainted by operating under modern, not historical, elo values.

Last but not the least, consider that under current algorithm, if zk would grow 10-fold, you would get paired with better people the better you play. The ideal solution is still procurement of an infinite playerbase, which would allow a 1600 elo player to play 10v10 where each other player is 1600.