[Search] Search   [Recent Topics] Recent   [Hottest Topics] Hottest   [Login] Login
modify ELO rating based on map
Forum Index » New Feature Request
atest

Messages: 9,
Joined: Feb 19, 2016,
Offline

atest

Messages: 9,
Joined: Feb 19, 2016,
Offline
Motivation
Some maps are unbalanced and will at high level play lead to near 100% win percentage for certain race combinations. Intuitively this should balance out in the long run as you get the same amount of good and bad maps, this is however not true for the top 100 players, as a loss to a low rated player on an unfair map looses you a lot of points and there are now high rated players you can steal points so the reverse never happens.

suggestion
Record the number of wins and losses for all 9 race combinations (kha vs sap is different from sap vs kha) on all maps and use this modified formular:

Ea = Expected score player A
Eb = Expected score player B
Ra = true strenght player A (current points)
Rb = true strenght player B (current points)
Ma = Chance that player A wins on this map
Mb = Chance that player B wins on this map (1 - Ma)

Ea = 64Mb/(1+10^((Rb-Ra)/400))
Eb = 64Ma/(1+10^((Ra-Rb)/400))

Please note that if the map is fair Ma = Mb = 0.5 then the formular defaults to the current formular

Ea = 32/(1+10^((Rb-Ra)/400))
Eb = 32/(1+10^((Ra-Rb)/400))


limitations / Suggestions
maybe cap the map win/loss to 10% and 90%
use pseudo counts (ie 5 fake wins and losses for all combinations) to make it work on less popular maps
only count ranked random games as other games might be biased by one player playing the same map over and over again


The formula:
Ea = Expected score player A
Eb = Expected score player B
Ra = true strenght player A (current points)
Rb = true strenght player B (current points)

Ea = 1/(1+10^((Rb-Ra)/400))
Eb = 1/(1+10^((Ra-Rb)/400))
[Avatar]
simsverd

Messages: 921,
Joined: Dec 02, 2010,
Location: Bergen, Norway
Offline

[Avatar]
simsverd

Messages: 921,
Joined: Dec 02, 2010,
Location: Bergen, Norway
Offline
The problem with using data from win/loss is that you will have to use data from only the best players., thus gathering enough data for this to work properly will be an issue on most maps.

The whole point raised by other top players also, is to make a better point distribution system for the top of the ladder - mainly to give the top players an incentive to play (more) rated random games.

As it is now they often get paired with players with 2-300 less points wich will deliver points in the edges of what ELO formula gives (1-3 points if they win and 30-32 points loss if they lose).

There are currently some things in the works that will affect this:
- weighted vote, giving more influence to better players - this will probably remove some of the unfair maps over time
- new pro solo ladder with mirror games. This will negate the effect of bad maps
but these are not enough...

the most important 2 things to improve:
1. There should be a limit to how little ammount of points that can be won in random.. for example 8 points
2. Reduce the impact of a loss on unfair map. This can for example be done by automatically lowering the K-factor in the ELO forumula if the difference in points is big (either based on %difference in points or perhaps to a set value - K=16 if points more than 200 in difference)


Moderator of gamechat and forum. Tourney admin.

Send me a PM here or invite me to a game if you want to ask me something, suggest a map for tourney or just wanna play a game
[Email]
[Avatar]
xavi

Messages: 554,
Joined: Jun 12, 2009,
Location: California
Offline

[Avatar]
xavi

Messages: 554,
Joined: Jun 12, 2009,
Location: California
Offline
We are looking into improving this very soon. Hopefully, this week or the next one. No decision has been taken so far but I will comment further on this thread once we have reached a conclusion.

Some players also recommended to use the ECF grading system
https://en.wikipedia.org/wiki/ECF_grading_system

-= The UniWar Team =-
[Email]
atest

Messages: 9,
Joined: Feb 19, 2016,
Offline

atest

Messages: 9,
Joined: Feb 19, 2016,
Offline
Sorry for the late reply

The problem with using data from win/loss is that you will have to use data from only the best players., thus gathering enough data for this to work properly will be an issue on most maps.


While this is technically true, even if noobs win 60/40 and 2000+ rated players win 90/10 on a map, my formula would still be more fair since 60/40 is closer to 90/10 than 50/50


There are currently some things in the works that will affect this:
- weighted vote, giving more influence to better players - this will probably remove some of the unfair maps over time
- new pro solo ladder with mirror games. This will negate the effect of bad maps
but these are not enough...


I don't understand why mirror will change the bad maps, if a map is unfair and you play it mirrored the you are playing two unbalanced games one which looses you 30 points and one which gains you 3.
I like the weighted vote idear, but are we sure that people only vote based on balance. if a map has 30% down votes is that then because players find it unbalanced or boring?


the most important 2 things to improve:
1. There should be a limit to how little ammount of points that can be won in random.. for example 8 points
2. Reduce the impact of a loss on unfair map. This can for example be done by automatically lowering the K-factor in the ELO forumula if the difference in points is big (either based on %difference in points or perhaps to a set value - K=16 if points more than 200 in difference)


if you always win 8 points then everybody can get infinite score by playing against bots, they also get an infinite score against everybody they are more than 80% likely to beat.
if you change the K factor when the difference is more than 200 then a player like Earth will get half points vs all but 15 players and now he will loose 16 and gain 1-2 instead of loosing 32 and gaining 3... so I don't see how that changes anything

Some players also recommended to use the ECF grading system
https://en.wikipedia.org/wiki/ECF_grading_system


I think that is a good idear, but appart from "creating points" if active players win more than inactive players i don't see how it is much different from ELO...It has the arbitrary score floor... but I don't think that is the correct solution, because people don't complain that they lose every now and then due to RNG, they complain they loose due to maps.
Forum Index » New Feature Request
Powered by JForum 2.1.9 © - 2020-04-14 v124 - UniWar website