every dota 2 player should have equal win rate?

[bloodchow]

every dota player should stay close to 50% win/lose rate as much as possible but their rating is maintained based on their level

lets say :

player A is a good dota player but his win rate is 55% and his actual rating is around 1700+

player B is less skilled than player A but his win rate is 44% and his actual rating is around 1300+

so we need player A to team with less skilled players in order to lose some games to maintain to 50% win rate while staying around 1700 rating

player B should be put with better players to maintain 50% win rate while staying around 1300 rating

so conclusion is player A is harder to win games while player B is easier to win games to maintain 50% win/lose rate

[Mizure]

Originally Posted by bloodchow

every dota player should stay close to 50% win/lose rate as much as possible but their rating is maintained based on their level

lets say :

player A is a good dota player but his win rate is 55% and his actual rating is around 1700+

player B is less skilled than player A but his win rate is 44% and his actual rating is around 1300+

so we need player A to team with less skilled players in order to lose some games to maintain to 50% win rate while staying around 1700 rating

player B should be put with better players to maintain 50% win rate while staying around 1300 rating

so conclusion is player A is harder to win games while player B is easier to win games to maintain 50% win/lose rate

iam 1989.

but who could possible be 1300/1700 ?

[NoThlnG]

Well, if every game you have 50% to win or lose it means both team is balanced which means the game is fun.

So, where's the problem here?

[Simon493]

Relative rating systems can really only be asymptotically efficient. You can design something where you converge in probability on a 50% win rate but very high/very low skill people will always take longer to get there.

[JippyTheLion]

It seems more complicated than that. In the situation you described, it seems like the actual skill of player B could be lower than 1300. In this case, it sounds like that player ended up with too high of a rating, and ended up playing a number of teams that were better an ended up with a win rate below 50%. Eventually he will lose enough games and enough rating to match him up with people that he can consistently win against, and the rating would go up again. Eventually this oscillation above and below 50% win rate would settle to 50%, assuming the skill of the player stays constant. The opposite is also true for player A.

A logical way for the matchmaking system to work would be to continue matching players based on the matchmaking, without considering the win rates. Eventually, the players would gain or lose enough points to put them at a level where they maintain a 50% win rate.

Of course, Dota is also a team game, which makes all of the matchmaking much more difficult to do. A team that just doesn't work well together could cause some unjust loses, and vice-versa.

In addition to all of this, there will also be no way to make players at very high or very low rankings have a 50% win rate. There is simply no way to create even teams for those players at the level they are at.

[Simon493]

Originally Posted by JippyTheLion

A logical way for the matchmaking system to work would be to continue matching players based on the matchmaking, without considering the win rates. Eventually, the players would gain or lose enough points to put them at a level where they maintain a 50% win rate.

You have the same problem really. When someone loses you really only gain the information that their team were worse than whoever they played against and given the relatively high variance in performance given a certain skill level 1 win or loss doesn't actually contain that much useful information about a players skill level. Assuming skill is normally distributed(or any really) as you move away from the mean it will take longer and longer for people to hit their "true" rating.

But rating systems like in HON do have the advantage that you can have unequally skilled teams and weight points won/lost accordingly. Either way there is no way to get around the fact they are only asymptotically efficient.

You can do things like have points scale faster when you first start playing, have consistently high kdr or many wins in a row but your sort of trading off the efficiency with the speed of convergence. (SC2 uses a similar system with their placement system and use your 5 games to do a rough estimate of your skill level. But estimating skill is FAR FAR easier in a game like sc2 where its 1v1).

[bloodchow]

bump need more ideas from everyone

[Simon493]

I'm sure valve has professional statisticians to work on things like this. From what I've seen come out of their comments about steam sales I know they have at least one professional econometrician. If your personally interested in rating systems the elo would be the place to start Elo rating system - Wikipedia, the free encyclopedia. I havn't really read a huge amount on it and it is for 1v1 (chess) but again because of all the different things in dota2 that effect your chances of winning (namely the 4 other players) it would really only hold asymptotically and you wouldn't want to use as large of a "k" factor (rate at which ratings change) or have a variable k factor w.r.t something like games played or skill level.

[venze]

Sorry for off topic, but i have little problem in beta - my lvl doesnt grow up at all, but i have about 6 wins \ 1 lose. Is this because of friend gave me a gift-based invite or smth else? Thank ya

[JippyTheLion]

Originally Posted by Simon493

You have the same problem really. When someone loses you really only gain the information that their team were worse than whoever they played against and given the relatively high variance in performance given a certain skill level 1 win or loss doesn't actually contain that much useful information about a players skill level. Assuming skill is normally distributed(or any really) as you move away from the mean it will take longer and longer for people to hit their "true" rating.

But rating systems like in HON do have the advantage that you can have unequally skilled teams and weight points won/lost accordingly. Either way there is no way to get around the fact they are only asymptotically efficient.

You can do things like have points scale faster when you first start playing, have consistently high kdr or many wins in a row but your sort of trading off the efficiency with the speed of convergence. (SC2 uses a similar system with their placement system and use your 5 games to do a rough estimate of your skill level. But estimating skill is FAR FAR easier in a game like sc2 where its 1v1).

It seems that you've caught on that my main experience is in Starcraft 2, where team game balance is terrible and expected to be terrible =P
You are definitely correct that it is far easier to balance a 1v1 game, and even then it can take a frustratingly long time for the skill level to converge. Something important to note is that 5 games to place a new player is nowhere near enough. While a large portion of the community is placed correctly, there are still, many, many, players who end up in a much higher or lower league than they should be in.

It seems like no one has really figured out a good way to balance a 5v5 matchmaking system. I have not played HoN, and am not familiar with the system used there, but from what you describe maybe something like that would work. Adding additional stats into the pool of things considered would also probably help, but then you still might end up with some of the problems you see in Starcraft.

In general it seems like you know a lot more than I do on the subject. My expertise unfortunately only covers Starcraft 2.

[LEGEND OF DIV]

Originally Posted by bloodchow

every dota player should stay close to 50% win/lose rate as much as possible but their rating is maintained based on their level

lets say :

player A is a good dota player but his win rate is 55% and his actual rating is around 1700+

player B is less skilled than player A but his win rate is 44% and his actual rating is around 1300+

so we need player A to team with less skilled players in order to lose some games to maintain to 50% win rate while staying around 1700 rating

player B should be put with better players to maintain 50% win rate while staying around 1300 rating

so conclusion is player A is harder to win games while player B is easier to win games to maintain 50% win/lose rate

ur correct in every working balance system like div-league alle players end up 50% winrate (this is forced)

[bloodchow]

Originally Posted by LEGEND OF DIV

ur correct in every working balance system like div-league alle players end up 50% winrate (this is forced)

better dota 2 implement this system, because i saw good hon player always have high rating but win rate is way more than 50%

[Simon493]

Well your never going to get anything that is perfect off of 5 games, the sample size is just too low. Even someone completely horrible could win 5 in a row if they got lucky enough. There simply isn't enough information contained in the 5 games to judge a players skill.

The idea of a rating converging in probability is that as the number of games played increases to infinity your rating converges to a "true" rating. In this case it would in reality be a range because your skill level varies from game to game(don't play hon/dota high, trust me).

Now you can have ratings converge faster by using more information but it also makes them less accurate.

Take KDR for example.(Kill death ratio)

If we add it into our rating we will see people converge towards their true rating quicker. Think of it like this, whats more likely:

a very low skilled person to win 5 games

or

a very low skilled person to win 5 games AND get a very high KDR.

Obviously the first is far more plausible.

HOWEVER when we introduce KDR we introduce model misspecification. Presumably we define "skill" as the ability to win games. Now while KDR is likely highly correlated with "skill" (especially at the start before the rating system starts to separate high/low skill people) it doesn't necessarily hold for all people.

Consider two players who play support. The first steals kills from carries. The second sets up kills for the carries. Given the same "skill" level (ability to win games) the first will be rated higher if we include KDR in the ratings, despite being a worse player. Also it adds in a negative behavioral incentives (kill stealing/picking gankers/carries and so on).

[bloodchow]

Originally Posted by Simon493

Well your never going to get anything that is perfect off of 5 games, the sample size is just too low. Even someone completely horrible could win 5 in a row if they got lucky enough. There simply isn't enough information contained in the 5 games to judge a players skill.

The idea of a rating converging in probability is that as the number of games played increases to infinity your rating converges to a "true" rating. In this case it would in reality be a range because your skill level varies from game to game(don't play hon/dota high, trust me).

Now you can have ratings converge faster by using more information but it also makes them less accurate.

Take KDR for example.(Kill death ratio)

If we add it into our rating we will see people converge towards their true rating quicker. Think of it like this, whats more likely:

a very low skilled person to win 5 games

or

a very low skilled person to win 5 games AND get a very high KDR.

Obviously the first is far more plausible.

HOWEVER when we introduce KDR we introduce model misspecification. Presumably we define "skill" as the ability to win games. Now while KDR is likely highly correlated with "skill" (especially at the start before the rating system starts to separate high/low skill people) it doesn't necessarily hold for all people.

Consider two players who play support. The first steals kills from carries. The second sets up kills for the carries. Given the same "skill" level (ability to win games) the first will be rated higher, despite being a worse player. Also it adds in a negative behavioral incentives (kill stealing/picking gankers/carries and so on).

those supporters go 0-10 in dota 2 tournament on august means they are sucks?

[Simon493]

No, quite the opposite really.

[benybc]

i dont know if this is right

[JippyTheLion]

Originally Posted by Simon493

Well your never going to get anything that is perfect off of 5 games, the sample size is just too low. Even someone completely horrible could win 5 in a row if they got lucky enough. There simply isn't enough information contained in the 5 games to judge a players skill.

The idea of a rating converging in probability is that as the number of games played increases to infinity your rating converges to a "true" rating. In this case it would in reality be a range because your skill level varies from game to game(don't play hon/dota high, trust me).

Now you can have ratings converge faster by using more information but it also makes them less accurate.

Take KDR for example.(Kill death ratio)

If we add it into our rating we will see people converge towards their true rating quicker. Think of it like this, whats more likely:

a very low skilled person to win 5 games

or

a very low skilled person to win 5 games AND get a very high KDR.

Obviously the first is far more plausible.

HOWEVER when we introduce KDR we introduce model misspecification. Presumably we define "skill" as the ability to win games. Now while KDR is likely highly correlated with "skill" (especially at the start before the rating system starts to separate high/low skill people) it doesn't necessarily hold for all people.

Consider two players who play support. The first steals kills from carries. The second sets up kills for the carries. Given the same "skill" level (ability to win games) the first will be rated higher if we include KDR in the ratings, despite being a worse player. Also it adds in a negative behavioral incentives (kill stealing/picking gankers/carries and so on).

Ah, I hadn't even consider the possibility of encouraging KS-ing or other bad habits. Those are probably prevalent enough as it is.

With that in mind, you could define different benefits or bonuses based on the hero being played, but that seems like a very slippery slope to start down, and only gets more complicated as you expand it to more and more heros. Personally, I would rather have it take more time to converge than have the skill level converge quickly to an inaccurate value, but there might be room for a hybrid system of sorts. Use bonuses for the first few games to get a general guess, and then use a win/loss based system after that.

It seems that I've made a circle somewhere and gone right back to one of the ideas in your first post. Silly me. I should probably leave the speculating to those who know more about it.

[Simon493]

Originally Posted by LEGEND OF DIV

ur correct in every working balance system like div-league alle players end up 50% winrate (this is forced)

You can't really force this without significantly reducing the efficiency of your rating system.

When you force a 50% win rate(you can't actually force a 50% but I'll be charitable and assume a roughly 50% win rate) you are essentially saying that peoples ratings change at very extreme rates or you have uneven games. Intuitively this makes sense as if someone keeps winning you need them to lose (face much harder opponents) to maintain a near 50% win rate.

The problem with this is that each win/loss doesn't actually contain that much information about a players skill. IE it is completely plausible for a player who is terrible to win his first 5 games just because he is lucky. To force a ~50% win rate you would need this persons "rating" to increase VERY VERY quickly as they win successive games and all the sudden you have these low-skill players in higher skilled games just by random chance.

There is no way to develop a PERFECT system that works for 100% of players 100% of the time. But a system like in hon has a ~50% win rate for ~95% of the players. Sure you get the outliers who have very high and very low win % as they move up or down to their true rating but that is a somewhat unavoidable problem.

You can use a few different things to reduce this but really there is no perfect way to do it and it will pretty much always be a problem. There is no way for the system to differentiate between a "1600"(slightly above average) player from a "1900"(very high skilled) player if they both start out with 10 wins and 0 losses. If you having something like a forced ~50% win rate people are going to be flying up and down the rating system as they go on streaks that are more or less luck based or you have force very uneven games (ie you get "lucky" and win too many in a row and get matched up against opponents with much higher skill levels).

[Simon493]

Originally Posted by JippyTheLion

Ah, I hadn't even consider the possibility of encouraging KS-ing or other bad habits. Those are probably prevalent enough as it is.

With that in mind, you could define different benefits or bonuses based on the hero being played, but that seems like a very slippery slope to start down, and only gets more complicated as you expand it to more and more heros. Personally, I would rather have it take more time to converge than have the skill level converge quickly to an inaccurate value, but there might be room for a hybrid system of sorts. Use bonuses for the first few games to get a general guess, and then use a win/loss based system after that.

It seems that I've made a circle somewhere and gone right back to one of the ideas in your first post. Silly me. I should probably leave the speculating to those who know more about it.

AFAIK that is actually what SC2 does (using bonuses for first couple of games) because I got placed in silver despite losing every single one of my placement matches But yeah that is probably one of the better ways to do it. KDR relative to team average as some sort of multiplicative bonus on rating gain/loss for first 10-20 games. If they wanted to they could use data from beta to construct a better system. Take a look at the first 10-20 games of every player and run regressions and see what separates high/medium/low skill players then have those as multiplicative bonuses for first x games. (But again we wouldn't want to include it in our long term estimate of skill because even though it might be highly correlated with true skill it likely doesn't necessarily hold for every player, like we saw with KDR)

[StudyMan]

you are 1500 rating, if you play with people under your skill, you will win increasing ur stats.

if you are 1500 rating and you play with people above your skill you will lose.

if you play with people of the same skill, you are gonna win/lose equally, or else your rating will change into an area that you can't win/can't loose.

so ultimately, people with 1700 rating will get there, and people with 1300 rating will get there, and they will stay around that rating, so people will end up with 50% W/L eventually.

it's not the egg that laid the chicken op, it's the other way around.