]]>
santori wrote:
Can Sheeo please comment on how the code ACTUALLY works and the reasoning behind it since he is the person who actually wrote it (at least in its current incarnation)? This discussion is pretty speculative from all parties and I think that it would a lot more productive if it were more grounded in fact.
Can Sheeo please comment on how the code ACTUALLY works and the reasoning behind it since he is the person who actually wrote it (at least in its current incarnation)? This discussion is pretty speculative from all parties and I think that it would a lot more productive if it were more grounded in fact.
In case you ever need to reach me or need me to reply to something, feel free to pm me here on the forums or at [email protected]
Statistics: Posted by Sheeo — 29 Oct 2015, 04:29
]]>
santori wrote:
I realize that making code change suggestions when you haven't tested the code yourself is a pretty unreliable and there are maybe other implications of changing the game quality floor to a nonzero value. However, I think that this is something that could probably be changed without a huge amount of work.
Maybe the code should be changed in another way - but I feel like we should try to do something to prevent terrible matchups.
Please let me know if my interpretation of the code is incorrect or my armchair analysis is nonsensical.
Thanks,
santori
I realize that making code change suggestions when you haven't tested the code yourself is a pretty unreliable and there are maybe other implications of changing the game quality floor to a nonzero value. However, I think that this is something that could probably be changed without a huge amount of work.
Maybe the code should be changed in another way - but I feel like we should try to do something to prevent terrible matchups.
Please let me know if my interpretation of the code is incorrect or my armchair analysis is nonsensical.
Thanks,
santori
Hey santori,
I hadn't seen these posts until now, apologies.
Nice of you to look into this code
While the currently running code isn't exactly what you pasted, the matching functionality should be equivalent to the current production code (Check out the master branch, where the original author is not myself).santori wrote:
So maybe changing:
- Code:
self._deviation_quality = {
450: 0.01,
350: 0.1,
300: 0.7,
250: 0.75,
0: 0.8
}
to something like
- Code:
self._deviation_quality = {
450: 0.4,
350: 0.5,
300: 0.7,
250: 0.75,
0: 0.8
}
A new player gets the rating (1500, 500) initially. Matching with a highly rated ladder veteran yields a match quality of 14%:
- Code:
>>> trueskill.quality_1vs1(Rating(1500, 500), Rating(2515, 48))
0.14695718222057164
And with an average joe 51%:
- Code:
>>> trueskill.quality_1vs1(Rating(1500, 500), Rating(1200, 60))
0.5102052390139458
It's fair to say that the threshold for dev>450 make no sense, since it's practically impossible to gain a 1% match quality with any other player in the ladder. Quality vs a theoretical (3000, 60) rated player would be 3%.
In practice, a deviation of 450+ means you haven't played any games, since the first game even against another (1500, 500) rated player will give you a deviation of <450.
On this basis I can't disagree with you and think we should definitely raise the threshold to a less extreme value. I like 40%.
My only concern is that our ladder game throughput is ~26 games per hour, and we do want to make sure players actually get games played. But I don't think raising this will cause less matchups.
Statistics: Posted by Sheeo — 29 Oct 2015, 04:27
]]>
Statistics: Posted by Gorton — 27 Oct 2015, 19:23
]]>
]]>
0's in our community get kicked from global games, get crushed in ladder, and end up with negative rating in both ladder and global
I used to be second worst on the ladder. Then, I reached out to a trainer and got tips and instantly rocketed to 1000 (the point at which you start learning by yourself. I was still going up but at a slower pace (due to not too many people being above 1400).
I think the main problem is that the 0 rated people do not know trainers exist and/or don't ask anyone. I know that I had to dig out the forum thread and then wait for one of the people to log in.
Whenever I scroll down the online users list I don't see many trainer avatars (very often none). Either the trainer team is too inactive or no one wears the avatar.
Statistics: Posted by TheKoopa — 27 Oct 2015, 18:47
]]>
]]>
Even if it's just 1200 or 1300 rating, many of those players don't play ladder or hardly play it. There was a thread on "what do you play" a while ago, with spraedsheets and all. Would be nice to compare the numbers. In any case I'm sure you will agree that ladder is played way less than custom, and high ladder rating is harder to get (for that reason) as well.
Statistics: Posted by Col_Walter_Kurtz — 27 Oct 2015, 13:45
]]>
Col_Walter_Kurtz wrote:
It's not a fairy tale at all. People play far more custom games. 1500+ custom rating and 0 ladder rating happens all the time. Vice versa would be a rarity. The exact number of games needed to get to an accurate ladder rating is uncertain, but it's definitely more than 10, or the handful of games that unbalanced match ups may take place for newer players.
santori wrote:Once again, I think that the 1500 level joining FAF for the first time is a fairy tale. Even if it isn't, it will take much fewer than 100 games for a player to get the correct rating. Trueskill would be a pretty horrible rating system if it took that long.
It's not a fairy tale at all. People play far more custom games. 1500+ custom rating and 0 ladder rating happens all the time. Vice versa would be a rarity. The exact number of games needed to get to an accurate ladder rating is uncertain, but it's definitely more than 10, or the handful of games that unbalanced match ups may take place for newer players.
This can happen, but you are exaggerating tons. It's extremely rare. It's even rare for a player to reach 1500+ custom rating (we maybe get 1 per week or something) and then add on top of that likelihood of them not having played ladder before... it's pretty unlikely. Practically irrelevant.
E: actually 1 per week is very optimistic, probably more like 1 per month if you exclude players who have previously dropped out of 1500
Statistics: Posted by Blodir — 27 Oct 2015, 12:29
]]>
santori wrote:
Once again, I think that the 1500 level joining FAF for the first time is a fairy tale. Even if it isn't, it will take much fewer than 100 games for a player to get the correct rating. Trueskill would be a pretty horrible rating system if it took that long.
Once again, I think that the 1500 level joining FAF for the first time is a fairy tale. Even if it isn't, it will take much fewer than 100 games for a player to get the correct rating. Trueskill would be a pretty horrible rating system if it took that long.
It's not a fairy tale at all. People play far more custom games. 1500+ custom rating and 0 ladder rating happens all the time. Vice versa would be a rarity. The exact number of games needed to get to an accurate ladder rating is uncertain, but it's definitely more than 10, or the handful of games that unbalanced match ups may take place for newer players.
Statistics: Posted by Col_Walter_Kurtz — 27 Oct 2015, 12:11
]]>
Gorton wrote:
As far as I am aware, a while ago it was made impossible that you can play people above 1400 or so 1v1 rating. Not sure on specifics.
As far as I am aware, a while ago it was made impossible that you can play people above 1400 or so 1v1 rating. Not sure on specifics.
This may have been implemented previously in the server code but right now the only thing that is taken into account is game quality. Since game qualities of 0 percent are allowed, anyone can play anyone.
Reaper Zwei wrote:
Personally I like Blodir's idea of limiting the search range. It's not all that much fun when you know who's going to win right from the get go.
I like Blodir's idea of limiting the search range also. That is in effect what would happen if there was a minimum allowed game quality. If the rating gap for a proposed match was too wide (like 0 rated vs 1000 rated) then the game quality would be too low to be allowed and the matchup wouldn't happen.
madformuse wrote:
On a personal level when I started playing a year ago it was pretty intimidating facing off against stronger players, the warnings helped but I would have rather started out with some 0 rated opponents. The concern then is, not that new players can't read but they don't like what they see.
I may not be representative though, when my friend started playing it didn't bother him at all.
I agree with your sentiment. And given that many new players expressed to me that they are unhappy that a severe mismatch has happened, I think that is a common feeling.
Anihilnine wrote:
Well when i was 1399 i played a lot of zero rated players and
a) it was annoying for me to play against them ... to make it interesting i went super early t2 or used labs instead of tanks etc.
b) they were very unhappy, calling it bullshit
c) lots of insta-ctrl-ks
I'm sure there is a reason why the 15-game-or-whatever rule exists but if there are viable alternatives lets try some of them. This way sucks
Yeah. I"ve seen the same thing.
There are two parts of every player's 1v1 score - mean and deviation. The mean is the score that you see in game and in the ladder rankings. The deviation is basically an estimate of how *accurate* the system thinks that your mean score is. You don't see this part of your score in your 1v1 rankings but it is stored in the database. The lower this number is, the more accurate your mean ranking is. I think that the "15 games" is an estimate of how many games you need to play before this number gets low. Once the number is low enough, the system will stop instantly giving you very low quality games (i.e. rating 0 vs 1000) type games.
Anihilnine wrote:
I think its because when the new player I awesome, they win games against 1500 rated players, which does them by a huge amount. The guessing period exists so those players don't unfairly lose rating
I think that's the logic that it follows as well. However, I really really doubt that anyone new to multiplayer Forged Alliance has EVER started beating 1500 rated players right away.
Col_Walter_Kurtz wrote:
The way I see it, when a good player starts at 0 ladder rating it will take 100 games until he is 1500 rated. Se he's going to annoy the sub 1500 players for a 100 games in a row. So either new players get the chance to play random opponents like it is now or every time a capable player joins ladder the lower rated players will get him in rotation for much longer.
This thread is a bit unnecessary. Your first 10 ladder games can be over in an hour, and you will get players of your skill level too.
Once again, I think that the 1500 level joining FAF for the first time is a fairy tale. Even if it isn't, it will take much fewer than 100 games for a player to get the correct rating. Trueskill would be a pretty horrible rating system if it took that long.
Aurion wrote:
It's just not smart from a 'we want new players to play ladder' point of view. I handled the random period without issues, but many people will get discouraged. They won't play ten games if they get totally stomped in the first two.
Yeah this is exactly the scenario that we should strive to avoid if we want the community to be healthy long term.
Can Sheeo please comment on how the code ACTUALLY works and the reasoning behind it since he is the person who actually wrote it (at least in its current incarnation)? This discussion is pretty speculative from all parties and I think that it would a lot more productive if it were more grounded in fact.
Statistics: Posted by Eric_Lesch — 27 Oct 2015, 08:42
]]>
Col_Walter_Kurtz wrote:
The way I see it, when a good player starts at 0 ladder rating it will take 100 games until he is 1500 rated. Se he's going to annoy the sub 1500 players for a 100 games in a row. So either new players get the chance to play random opponents like it is now or every time a capable player joins ladder the lower rated players will get him in rotation for much longer.
This thread is a bit unnecessary. Your first 10 ladder games can be over in an hour, and you will get players of your skill level too.
The way I see it, when a good player starts at 0 ladder rating it will take 100 games until he is 1500 rated. Se he's going to annoy the sub 1500 players for a 100 games in a row. So either new players get the chance to play random opponents like it is now or every time a capable player joins ladder the lower rated players will get him in rotation for much longer.
This thread is a bit unnecessary. Your first 10 ladder games can be over in an hour, and you will get players of your skill level too.
It's just not smart from a 'we want new players to play ladder' point of view. I handled the random period without issues, but many people will get discouraged. They won't play ten games if they get totally stomped in the first two.
You could also get a middle ground as starting rating with a really high deviation when people start (500 for example). It really won't take a good player 100 games to reach 1500 by the way, after my first game I was at least 600. If a good player annoys X rating range for a long time, that means he's losing on purpose every now and then.
Statistics: Posted by Aurion — 26 Oct 2015, 18:00
]]>
This thread is a bit unnecessary. Your first 10 ladder games can be over in an hour, and you will get players of your skill level too.
Statistics: Posted by Col_Walter_Kurtz — 26 Oct 2015, 16:18
]]>
]]>
]]>