Elo Formula

[Gothmog]

Just work it on an experience system. You get 12 experience for killing an orc, and 35 for killing a hvy int. Level 30 should require something in the neighborhood of 500,000,000 experience.

[TheBored]

Okay, so first he tells us that we should kill Orcs. Maybe a typo, whatever. NOW hes telling us that we should get 14.2 MILLION kills against a HVY INT to get lv 30.

/glare.gif" style="vertical-align:middle" emoid=":glare:" border="0" alt="glare.gif" />

TB

[Mastametz]

elo needs to understand

stack early > stack late

[Psychosis]

not always true mastametz, if a team can hold their own decently until they get tech up, then gets a stack, they can win handily, and ELO wont even show that they had the stack.

[Gothmog]

QUOTE (_SRM_TheBored @ Sep 21 2006, 06:53 AM) Okay, so first he tells us that we should kill Orcs. Maybe a typo, whatever. NOW hes telling us that we should get 14.2 MILLION kills against a HVY INT to get lv 30.

/glare.gif" style="vertical-align:middle" emoid=":glare:" border="0" alt="glare.gif" />

TB

That was not a typo. Do I seem like the type that would make any word affiliated with allegiance look like "orc"? And yes, 14.2 million kills seems about right. Despite the obvious sarcasm of my post altogether, we'll pretend that hypothetically the point system I proposed was in existance. Level 30 would be leet, and require months upon months of killing and base raping and scouting and probing and nanning, etc, etc... It IS the highest level you can get to, after all. There can be only one. Go eat an ostrich.

[TheBored]

If that is an attempt at humor, it is completely lost on me. There are two options here, you aren't funny or you are just dumb.

You can make the pick, doesnt matter to me.

TB

EDIT: and your necro posts are annoying.

[Adaven]

it gave me a chuckle. But then, I'm pretty easily amused.

[Aoreias]

Looking above at what Frag posted over a month ago, I think he really hit the nail on the head about the ELO problem with allegiance these days.

First off, it most certainly does encourage stacking.

First assumption to support this point: People care about their standings. Lets be honest, we're all competitive $#@!s or we wouldn't play this game anyway, we'd all give each other pats on the back and go play HuggyFluffyBears.

Second assumption: A really good vet is worth more than 2 average players to a team. How big an advantage of random players would you give someone so you could have Weed on your team? Some people are 'worth' 4-5 people.

On that note, I'm going to ignore stacking because of unequal comms (If stats count what's my incentive to fly for a bad comm). Yes, I know it's another major source of stacking. I'm probably more guilty than most people of abusing this one.

The way we're using ELO right now, we just average it out on the team. This is all fine and dandy, but by accepting a good vet, you can next accept a noobie and counteract the elo gain that your good vet gives you, (assuming ELO is working anyway). Do it enough, and you end up with a nicely stacked team padded with some newbies to lower the score enough to make the game count. End result is what we see today, stacking.

In that example I assume that people are trying to make the game count, so they accept newbies. What about a commander who doesn't care about elo? They'll just keep accepting the vets who are waiting to join his team, and the game ends up stacked and not counting. This second scenario happens a lot these days. An awful lot. I would wager that of primetime->late night games 75-80% of games end up not counting because of an excessive stack.

Want to know how to really gain ELO now? Play for the best commanders and ride on their coattails. I've got the second-best win loss ratio in the top 50 players (second to Storm_Strike), which is better than Weed, Aarm, and Frag.

I am not better than Weed, Aarm, or Frag, but I do know how to pick 'em.

[Aoreias]

EDIT: Killing off first panrt of this post on ELO-autobalancing. After writing up the below code and iterating the results a number of times, ELO with autobalancing eventually comes to a fair convergence at about 1000 games. Past that point, in my model, the number start getting skewed as the 'good' players accrue massive amounts of elo and the 'poor' players massively negative elo.

Code: Select all

int[100] players; //True skill value of 100 players.  We don't get to see this number.

//give every player a true skill value that we're trying to discover
for (i = 0; i < 100; i++)
{
    players[i] = randomGaussDistr(30,30) //Average skill 30, gaussian distribution of 30.  Minimum value 0.
}

int[100] guess;    //This will hold the ELO score of players

//for each player, give everyone a default ELO value.
for (i = 0; i < 100; i++)
{
    guess[i] = 30;
}

//begin gameplaying

//pick 20 players for a game

//Sort based on ELO highest first.

//For each player

    //If A has a higher combined ELO score, place on Team B.  Otherwise place on team A.

//simulate game by adding up true skill values on both teams.  Give winning team's elo a +1, losing team's elo a -1.

//iterate gameplaying until convergence

//see how long convergence takes and discover if this is feasible, and if convergence actually happens.

[tmc]

I propose the following changes to make ELO converge:

1. There should be no penalty from "bailing". The amount of ELO lost should be directly related to the amount of the game that the player played. If someone leaves a game after 5 minutes, and his team ends up losing after a 3-hour game, he should not lose much ELO as he had absolutely nothing to do with the loss. Similarly, imagine having a lone vet in a team of newbs. The vet, however, is so good, that while he's there, he manages to single-handedly give his team the advantage. He's about to single-handedly beat the enemy team when his wife comes home and he has to leave. His team proceeds to lose, and he is deducted ELO. This prevents convergence.

2. Having vets in your team early is more valuable than late in the game. ELO should be weighed against a time-decaying function in order to calculate the odds of winning for each team.

3. Commanders are more important than pilots, and their win% should be calculated into the team's ELO (maybe as a multiplicative coefficient; somewhere between 0.8-1.25 might work). This would require ELO to be able to know who the commander is and take care of commander changes midgame.

4. We need to Calculate an actual value for the parameter which assigns an expected value to the team's ELO. We have enough data to do this already; divide up the ELO differences we have, and plot them against their win%. Find the equation, and use that.



Think thats it for now.