<< Back to Ladder Forum   Search

Posts 1 - 1 of 1   
Problems of the current rating system (1v1 ladder): 3/11/2017 23:09:20

Level 63
First, I want to give a review, on why I think the current rating system does not work for the 1 vs. 1 ladder. For that reason, I will focus on the things, that don’t work. The 1 vs. 1 ladder does use the Bayesian Elo system (https://www.remi-coulom.fr/Bayesian-Elo/) with games expiring after 5 month.


1) Expiring games

Let’s imaging, MoD joins the ladder for a ladder run and rates 2500 Elo for a short period. To get there he plays about 3 month. After this 3 month, he beats Motoki, Timinator-apex and some players rated with 2200 Elo points. The loss against MoD does not hurt them that much, as he is rated 2500 Elo. But after a while MoD decides to focus on other parts of warlight and does leave the ladder. What happens now is, that his rating decreases and he rates 1400 Elo just before his games against Motoki and Timinator-apex expire. This lost game against a ‘1400 Elo player’ does hurt their rating a lot.
This happens all the time, when 2000 + players do beat players with 1800 and then leave the ladder for a new ladder run. It leads to underrated 1800 players, from whom other players (2000+) can only get less points.
To avoid this problem most top players (intentionally or not) do ladder runs. These ladder runs cause even more problems, as they work like a snowball effect.
=> The current rating system tends to produce ladder runs.

2) Equal weighted games
Every game within the 5 month is rated equal. This may sound far, but it causes problems.
Player X joins the ladder and has a true skill of 1700 Elo points. He does play a lot of games within 2 month. As his true skill is 1700 Elo points, he will sooner or later rank 1700 Elo point in the system. He plays about 30 games at this level. But as he plays, he gets better and suddenly has a true skill of 2000 Elo. But to increase his skill level he has 2 options: He waits till his games expire (5 month) or he plays a lot of games. To make up his 30 games at skill level 1700 and get a rating of 1900 he would need to play more than 60 games on the level of 2000 Elo.
=> weighting games equal only shows your average skill over the last 5 month
As we did see in problem 1), expiring games with a cut (5 month) is a bad idea. This doesn’t work. Games with have been played in long ago, should be weighted less than ones, which did end recently.
Example 2:
Month 1:
Player A (rating 1500, true skill 1500) plays player B (rating 1500, true skill 1400) 3 times, 2 wins for A, 1 win for B
Month 2:
Player A (rating 1500, true skill 1500) plays player B (rating 1500, true skill 1500) 4 times, 2 wins for A, 2 win for B
Month 3:
Player A (rating 1500, true skill 1500) plays player B (rating 1500, true skill 1600) 3 times, 1 wins for A, 2 win for B
We would want to know in month 3, that player B is currently better than player A. But with games weighted equal, we will never know.
=> players will create ALTs to show their true skill now.

As these problems, can’t be fixed within the current point system, I would advise to change the system to one of these 2 options:
1. system of the real-time ladder
2. "normal" elo*, starting with a fixed elo of 1500 and no expiring games

Personally, I would choose option 2.

* like it is described here (theory)
or here

I would take the most simple approach and start with a fixed value, as everything else, may produce more alts on the ladder.

Edited 3/11/2017 23:21:47
Posts 1 - 1 of 1