Edited by hughjaas at 11:26 CST, 25 November 2014 - 84753 Hits
Radiant its cool stuff. Reflex has it built in
But when challenged to explain how players can loose points on Qlranks while winning a game, the response was: Trueskill is better than Elo !!I literally quoted the parts of the paper that explain how that can happen.
THERE IS NO SET OF CIRCUMSTANCES in Trueskill where the winning player can loose score or RankingYes there is, I quoted the exact part of the paper that explains how it can happen.
But since you are arrogant SOBs, you couldn't admit that your multiplayer code is neither Elo nor Trueskill, and that it's not perfect.First of all, I have nothing to do with QLRanks. I asked Sirax after your last thread and he said they do use TrueSkill, so you're just wrong about that.
The TrueSkill skill of a player i is currently displayed as a conservative skill estimate given by the 1% lower quantile mu_i - 3*sigma_i. This choice ensures that the top of the leaderboards (a listing of all players according to mu - 3*sigma) are only populated by players that are highly skilled with high certainty, having worked up their way from 0 = mu_0 - 3*sigma_0.
...
If the skills are expected to vary over time, a Gaussian dynamics factor N (s_{i;t+1}; s_{i;t}; gamma^2) can be introduced which leads to an additive variance component of gamma^2 in the subsequent prior.
Edit: Here you go, I went to the trouble of finding it again:
The TrueSkill skill of a player i is currently displayed as a conservative skill estimate given by the 1% lower quantile mu_i - 3*sigma_i. This choice ensures that the top of the leaderboards (a listing of all players according to mu - 3*sigma) are only populated by players that are highly skilled with high certainty, having worked up their way from 0 = mu_0 - 3*sigma_0.
...
If the skills are expected to vary over time, a Gaussian dynamics factor N (s_{i;t+1}; s_{i;t}; gamma^2) can be introduced which leads to an additive variance component of gamma^2 in the subsequent prior.
Yes there is, I quoted the exact part of the paper that explains how it can happen.
I asked Sirax after your last thread and he said they do use TrueSkilCorrect, and I claim that sirax lied to you.
Honestly: I do not believe that you personally have a detailed understanding of Trueskill.Well I actually went and read the paper. Have you tried that? What kind of way is this to argue?
Should I draw you a fucking picture as well?
The system maintains a belief about player skill, represented by a probability distribution over all possible ratings. In order to condense this to a single value they use mu - 3*sigma, where mu is the mean and sigma is the standard deviation of the distribution. This is because it is more stable than the mean - it represents a lower bound on a players skill with 99% confidence.
As you play more games, it gets more confident in your skill estimate, so your standard deviation is lowered. This makes is difficult to raise your rating if you improve, so they make your standard deviation bigger before each calculation to compensate.
If you play a game that was too poorly matched, your true estimate (the mean) goes up by less than the additive factor on your standard deviation. So your displayed rating can go down.
One surprising thing is that if you have a really low standard deviation and play a game that has very bad match quality (see my accompanying paper for details but this usually means a very unfair match), it could be that your TrueSkill goes down after a win.http://www.moserware.com/2010/03/computing-your-skill.html
Try it, you can tweak Mu (mean) and sigma (standard deviation) any way you like, the winner CAN NOT loose ranking.
At which point moron2 points to the hyperreals and claims that I'm wrong (while this system includes infinitesimals).The hyper reals have infinitesmals, but not they don't allow you to divide by zero which is what I was arguing.