What would happen if chess used four tiers system instead of ELO
Let's say the lowest ELO for a (human) chess player is around 1250 and the highest ELO is around 2850. That makes difference between worst and best rated player = around 1600 ELO points.
As it is well known: 200 ELO difference (also refereed to as skill difference of a one "class") means that the weaker player should score about 25% of points while the stronger player should score about 75% of points. So that means out of 100 games
players with ELO 1400 is expected to win 25 games and losing 75 games against a player with ELO 1600. (we are ignoring draws for simplicity)
At 400 ELO difference the odds would be about 9 won vs. 91 lost game and so on.
So if chess used 4 tier system that would result splitting the playerbase in four large groups like this:
tier 1: ELO 1250 - 1650
tier 2: ELO 1650 - 2050
tier 3: ELO 2050 - 2450
tier 4: ELO 2450 - 2850
Where all players fitting into one particular group would be considered "about the same strength". That is at least what the QL tier system wants us to believe. However from the explanation above it should be clear that there would be a HUGE difference (400 ELO in worst case) between the worst and best players in one tier. It would also mean that in tier 3 we would have average club players together with international masters. Or in tier 4 we would have international masters together with the strongest of super grandmasters. And it should be obvious that none of this cases are even remotely "equal".
How does this translate to quake ? There may be differences between various game modes but basically I would expect the skill chain for quake (like the ELO difference between the best and the worst player) to be comparable to chess. You can speculate if this assumption is true or not but if you are willing to accept that skill chains in chess and quake are comparable than it also means that the main problem with tier system in QL is simply that splitting people into four groups - no matter how you do it - will always result in tiers where there can be up to 400 ELO difference between worst and best player in the group.
While for games to be "equal" you ideally need ELO difference around 100. Remember that Even 200 ELO difference in chess is considered a "class" difference in skill so even if skill chain for quake would be lower than in chess you would still end up with unbalanced tiers.
If you are asking why I am writing this it's because of two reasons:
1.) QL doesn't work again so I am killing time until it starts again
2.) It's my way how I try dealing with situation that QL devs opted for rating system which seems to be so trivially... let's say doubtful at best. And inherently flawed at worst.
I simply can't understand WHY ?
Thanks for attention.
Don't have time to go in-depth here but wanted to chime in and say I totally agree with you, and have been thinking the same.
Shouldn't have arbitrary non-overlapping adjacent discrete windows on the skill rating (as now). There is a particular problem when you look at the resulting distribution of skills that you will get gaussian clusters forming on the boundaries. Especially on the T3/T4 boundary which is around skillrating 75.
What do I mean by this? Well. When you jump into T4, you are suddenly faced with people many times better than you i.e. skill rating 250+, even 450+. On average the games will be >=200. Because of the way the score system is designed (I imagine they implement hacked ELO or some Bayesian system akin to Microsoft trueskill with factor graphs) your resulting score will be a function of your opponents much higher one and you will get kicked down hard, and the cycle will continue. To stay in Tier4 you actually have to be a very strong T4. This might be intentional but I don't like it.
They should remove the arbitrary skill discretisations and have a sliding window approach i.e. you are allowed to join servers with +=50 or 100 average of your current skill level (I do say average because we should encourage balanced mixed-skill scenarios on servers). Make all the servers totally skill-dynamic and then they can have less running. The sliding windows should grow with respect of skill if relevant i.e. I have a theory that real skill reaches diminished returns and increases in skill rating can be explained by situational/contextual advantage i.e. knowing the opponent, being good on a map. Any player 250+ could beat a 400 player sometimes. A 50 player would never beat a 100 player, ever.
Question -- how many people play ranked chess? And how many people play Quake Live?
Answer, as usual, is that there's not enough players to make the Elo or Tier system work much better. If anything, going down to 3 tiers is probably a better option than moving to 5.
Chess is also 1v1, and the QLRanks Elo system seems to be pretty adequate for duel. Skill ranking in team games is much trickier affair and requires a large and honest player base.
The windowed approach is interesting, but you'll be cutting off availability of games of the upper echelon of players, and so it's a non-starter.
Dr_Mr_Ed -- "but you'll be cutting off availability of games of the upper echelon" no I said increase the window size as a function of skill. "Skill ranking in team games is much trickier affair" not really, as I said they could use a Bayesian approach see page 22 of this document http://research.microsoft.com/en-us/...-wcci-2008.pdf
having discrete tiers is a total non-starter.
Yet another reason is that it messed up the stats, because all stats i.e. accuracy are a function of what tier you are in and they will suffer from the same discretisation on their distribution
"there's not enough players to make the Elo or Tier system work much better" this is nonsense, what made you draw that conclusion?
Last edited by Sasparillo; 02-05-2013 at 05:55 PM.
Under your system, all the 500+ players would essentially be lumped together in a single game. I don't know where you are, but except for weekend evenings, I don't see one of those players ever finding a match they'd be able to join.
And group ranking system needs a large variety of players to properly assess an individual's ranking. Think of it logically with a 10 player group versus 100 versus 1000. Which works better? But then break things down by region as we have to for ping reasons, and you splinter the base further. For good accuracy, you need a large and honest player base -- which we don't have.
@OP : There is a massive difference between QL and Chess : most of QL's gamemodes are team-based. Elo works fine on duel with QLranks' help and you can totally ignore QL's tiers for duel and only use QLRanks to choose your opponents. In fact, even in QL the server skills, as they appear to you (skill matched, more challenging...) aren't even a function of the server's tier, but of the difference between your skillrating and the server's (hence who is inside). Too bad QL's Elo implementation is too volatile (K too high) and has another flaw I'm not sure I can tell here. But for team modes life isn't just that easy.
@ecsplos1v : That sliding window approach was tried. It resulted in people often joining the easiest servers they could and progressively increasing the server's average skill in such a way that the original noob who joined the server could be facing a very strong opponent (along with the other intermediary ones). So yeah, it failed. Also, the rating system for non-duel gamemodes isn't a "hacked ELO or some Bayesian system akin to Microsoft trueskill with factor graphs". It's a very simple algorithm that I don't like much. Basically, it gives or takes away skillrating points according to your position on the final scoreboard once sorted by score/timePlayed. Top third gains points, bottom third loses points. It's pretty bad for very good player as they finish near the top of the scoreboard regardless of the team configurations and so, for them, the skillrating becomes a function of the number of games played rather than a function of their skill. But it works fine on normal players.
EDIT : Even if you take the "dynamic range" out of the tiers, you're stuck with static values and subject to problems similar to the current system. If your ranges are as wide as the tiers, you'll get pretty much the same results. If they are not, there may not be enough players in that range to be able to get a game going at all times of the day.
The solution, in my opinion, is to create as many tiers as we can according to the online players available and to a "minimum tier size". So if you have a minimum of 100 and 300 potential players are online, you can create 3 tiers of large width. If there are 10000, you can have 100 tiers of small width. The width is of course dependant on the skill... at the end of the curve the tiers will be much larger than in the middle. The more players are available, the better the skillmatching will be. When there are few players online, they will still be able to play because they won't be limited to a static, too small range : although the skill difference inside a server will be high, they'll be able to play. This system can of course have a different number of tiers per gamemode and region, according the each player's preferences. This is what I'm working on right now.
Last edited by FlashSoul; 02-02-2013 at 02:54 PM.
Are you implying that tiers are actually not used anywhere for matching the players ?
Originally Posted by FlashSoul
Yes. In duel, that is. Servers still have tiers though which can lead to weird situation where you cannot join a "More challenging" server because its tier is too low for you.
Originally Posted by tyhjyydesta
So in duels it is not used. But I meant if it is not used at all - including the most common game modes like FFA, CA, TDM, CTF. I was under the impression that it is being used there. But if it isn't used anywhere than what would be the purpose of such a tier system ?
It is used on non-duel gametypes. And its purpose is simply to separate a the players according to their skills. Sometimes too much, sometimes not enough.