Wesmania wrote:We actually had a whole discussion during dev conference regarding that (I believe it's posted somewhere on youtube). Changing starting rating for new players won't change anything - eventually this starting rating will become the new "average" as they play games against players with established rating. That's one of counter-arguments brought up on the conference.
That would ONLY be relevant if the current starting rating was actually the average rating!
WHICH IT IS CURRENTLY NOT! (in the dev discussion, you guys assume that it is. i have watched all of them)
Which is the whole problem, thats the whole thing that i am discussing since the second post in this thread! Thats how our True Skill system is SUPPOSED to be setup! There really should be nothing to discuss. The current starting rating needs to be the average rating, it is not, so it needs to be adjusted!
And yes, it is true, if we wait like 10 years (what you describe as "eventually"), until we have enough inflation accumulated to bring the average rating up to 1500, the problem has solved itself. Until then, all new players suffer. Why 10 years? Well, because in the last two years, our average rating has not yet become 1500. Far from it. So this "eventually" does not happen fast enough to correct this!
To make this clear:
I am not trying to adjust the starting rating AWAY from the average rating, im trying to adjust the starting rating TOWARDS the average rating faster than it would happen by itself!
Edit2:
Okay, i looked at the code. Looks like axle (the guy who implemented TrueSkill for FAF) simply set the average to 1500 (he could have set it to anything, because i guess at that time there was no rating system, right?) and then calculated the optimal TrueSkill parameters for that. So we need to change mu to whatever the current average is, then re-run his machine learning code to get the new optimal parameters, check if they make sense and if they are similar enough to the current parameters to not break stuff, and then make the changes to the parameters.
Then we need to do CTRL+F and search for "1500" in any server/api code and replace it with the new average, because it is, at least in the python server, hardcoded in a bazillion places (though most of them are tests). Doesn't sound too hard.
By the way, weren't people complaining that the rating fluctuates too much at some point? I imagine that things like that could happen if the actual rating average was different than the average that TrueSkill assumes, and that is currently the case.
Well, the current settings are:
(mu=1500, sigma=500, beta=240, tau=10, draw_probability=0.10)
We should recalculate axles stuff and see what we get with an updated mu. Or just follow the recommendations about how to set them. Beta and tau are mostly dependent on sigma anyways, and draw-propability we can simply calculate from games played. Or we can just leave them as they currently are, since they have been adjusted already by other people.
For anybody interested in what these parameters mean:
http://trueskill.org/beta – the distance which guarantees about 76% chance of winning. The recommended value is a half of sigma.
tau – the dynamic factor which restrains a fixation of rating. The recommended value is sigma per cent.