SLDB: Personal TrueSkill history graphs - Page 2

SLDB: Personal TrueSkill history graphs

Discuss development of lobby clients, server, autohosts and auto-download software.

Moderators: Moderators, Lobby Developers

User avatar
albator
Posts: 866
Joined: 14 Jan 2009, 14:20

Re: SLDB: Personal TrueSkill history graphs

Post by albator »

Two things:

1) I remember the hof page pointing me as 2nd till maybe October or so, so that maybe because the website does not use the same data.

2) More important, when checking the graph, I found out that from may till October, my ts goes from ~37+ till 35. And when I start to play again, the penalty points are removed (I assumed so since the website does not show any penalties), but the TS stays the same. It even decrease since I lost a game which makes sense. And since I thought the overall TS average of all player should stay 25, I thought I would have gain back the TS I lost through being inactive. My question is then: Has the TS been conservated and given to all the other active players so the average ts stay 25, or is it just definitely lost ? Something else ?
0 x

User avatar
Silentwings
Moderator
Posts: 3649
Joined: 25 Oct 2008, 00:23

Re: SLDB: Personal TrueSkill history graphs

Post by Silentwings »

And since I thought the overall TS average of all player should stay 25.
This is wrong - True Skill is not a "zero sum game" where you somehow trade your points with other players.

It's (basically) a Bayesian update scheme that learns the parameters of a Normal distribution representing the distribution of each players ability.
Has the TS been conservated and given to all the other active players so the average ts stay 25... ?
No.
0 x

User avatar
bibim
Lobby Developer
Posts: 913
Joined: 06 Dec 2007, 11:12

Re: SLDB: Personal TrueSkill history graphs

Post by bibim »

albator wrote:1) I remember the hof page pointing me as 2nd till maybe October or so, so that maybe because the website does not use the same data.
The website is supposed to use same data, however following points might explain the difference you observed:
  • The SLDB rating engine is entirely retroactive, which means it re-rates all needed past matches automatically when some accounts are merged/split (smurf detection/cancellation), as if it knew all the smurfs from the beginning. This alone can't justify what you observed, but it's just to explain that past data are recomputed from time to time, and thus can change.
  • I modified the SLDB rating engine recently to ignore the TeamFFA matches where teams are too uneven, in the same way it ignores Team matches too uneven (2v1 matches are ignored for example). This modification led to following TeamFFA games being ignored: 2v2v2v2v2v2v1v1 on Throne Acidic and 1v1v1v1v1v1v1v1v1v1v1v1v2v2v2v2 on Nation_v1. These matches were 2 TeamFFA games that you had won, so this change clearly impacted your TeamFFA rating negatively. I guess these 2 matches aren't so uneven after all and should actually be taken into account (2v2v2v2v2v2v1v1 isn't as unbalanced as 2v1 for example), so I think I will change a bit my uneven TeamFFA match detection.
albator wrote:from may till October, my ts goes from ~37+ till 35. And when I start to play again, the penalty points are removed (I assumed so since the website does not show any penalties), but the TS stays the same.
SLDB inactivity penalties don't work this way. They aren't "points" removed from a player and given back when he comes back. Inactivity penalties are a way to modelize the probable skill evolution of a player while he's not playing for a long time, through an increase of his skill uncertainty and a slight decrease of his estimated skill. As such, inactivity penalties can't be "removed", they just happen to your skill modelization while you're not playing.

I shouldn't even call them "inactivity penalties" which is a bit misleading, but I couldn't find better name ("iterations of inactive-player skill-evolution guess algorithm" might be more correct but not very usable). In a way, since these "penalties" increase your sigma (= skill uncertainty), you can consider them as an help to adjust your estimated skill faster to its true value when you come back (be it higher or lower than its current value). Playing matches when you come back will decrease back your sigma, so these "penalties" are actually removed by themselves by playing.

The inactivity penalty counter shown by SLDB and the replay site is decreased when the player has played enough games after coming back, but it's just an indicator, it doesn't artificially give you back any "point". When it's decreased back to zero, you can consider that the player has played enough games after coming back from his inactivity period for his skill to adapt back to its current true value and not be based on "guesses", so you can forget that he had received inactivity penalties.
albator wrote:And since I thought the overall TS average of all player should stay 25
What do you call "overall TS average of all player"? If you're simply speaking about the average of the estimated skills (mu values) of all the players, then the answer is no, and it can be easily shown with this example of 1v1 from Microsoft online trueskill rank calculator:

Image

The average of the estimated skills (mu values) is 25 before the game, but different from 25 after, because the mu value of the player having the highest sigma changes more than the other one.

Also, you can clearly see on this plot of the distribution of the mean skill levels for a popular Xbox Live game that the mean skill of the population isn't 25 in this case:
Image
0 x

User avatar
bibim
Lobby Developer
Posts: 913
Joined: 06 Dec 2007, 11:12

Re: SLDB: Personal TrueSkill history graphs

Post by bibim »

I changed the uneven TeamFFA games threshold, and launched a re-rate for BA since February. Here are your new TeamFFA ranks for current year:

Code: Select all

jan: 2nd
feb: 1st
mar: 1st
apr: 1st
may: 1st (6 inactivity penalties)
jun: 1st (+3 inactivity penalties)
jul: 1st (+8 inactivity penalties)
aug: 1st (+8 inactivity penalties)
sep: 2nd (+8 inactivity penalties)
oct: 2nd (+4 inactivity penalties)
nov: 2nd
dec: 2nd
0 x

User avatar
albator
Posts: 866
Joined: 14 Jan 2009, 14:20

Re: SLDB: Personal TrueSkill history graphs

Post by albator »

OK thanks ;) Was just trying to understand how it changed so rapidly.
0 x

User avatar
Jammer
Posts: 24
Joined: 16 Jun 2008, 11:37

Re: SLDB: Personal TrueSkill history graphs

Post by Jammer »

What can I do if the replay site isn't allowing me to log in with my lobby name and password? There doesn't seem to be any 'recovery' feature or contact information.
0 x

User avatar
Floris
Balanced Annihilation Developer
Posts: 606
Joined: 04 Jan 2011, 20:00

Re: SLDB: Personal TrueSkill history graphs

Post by Floris »

It will if you use the login button, and not the popup.
0 x

User avatar
Jammer
Posts: 24
Joined: 16 Jun 2008, 11:37

Re: SLDB: Personal TrueSkill history graphs

Post by Jammer »

I don't know a pop-up, just the 'login' button on the top right here:

http://replays.springrts.com/
0 x

dansan
Server Owner & Developer
Posts: 1198
Joined: 29 May 2010, 23:40

Re: SLDB: Personal TrueSkill history graphs

Post by dansan »

@Jammer: PM me your lobby username and the exact time and your timezone, then I can look into the logfile why you cannot login.
BTW: The login is case sensitive.
0 x

User avatar
Jammer
Posts: 24
Joined: 16 Jun 2008, 11:37

Re: SLDB: Personal TrueSkill history graphs

Post by Jammer »

did you get a chance to look at this? i sent a pm
0 x

dansan
Server Owner & Developer
Posts: 1198
Joined: 29 May 2010, 23:40

Re: SLDB: Personal TrueSkill history graphs

Post by dansan »

Hi,

sorry I was busy. I looked at it, and you're not the only one. Noone new to the site can login. Nothing I can do about it atm: It's a bug in Uberserver (the lobby server). I have informed the developers (https://github.com/spring/uberserver/issues/172). I'll inform here, once it's fixed.
0 x

dansan
Server Owner & Developer
Posts: 1198
Joined: 29 May 2010, 23:40

Re: SLDB: Personal TrueSkill history graphs

Post by dansan »

Abma fixed it - thank you!
→ Login works now.
0 x

raaar
Metal Factions Developer
Posts: 855
Joined: 20 Feb 2010, 12:17

Re: SLDB: Personal TrueSkill history graphs

Post by raaar »

I'm curious about how this actually works.

Recently springlobby started showing a trueskill column in the battle rooms.

Visible rank affects player behavior.

Do we have stats on how this works over time? This can encourage people to play...or inhibit them from playing, depending on the context.

So....if a player plays 10x against a better player and loses will it lose rank for each game or is there a cap?
0 x

dansan
Server Owner & Developer
Posts: 1198
Joined: 29 May 2010, 23:40

Re: SLDB: Personal TrueSkill history graphs

Post by dansan »

https://www.google.de/search?q=trueskill

Best article is IMO this one: http://www.moserware.com/2010/03/comput ... skill.html

There are graphs on you personal skill development on the replay site. They are generated by SLDB: http://replays.springrts.com/player/142479/
Others can see those graphs only if you have disabled privacy protection (like me: http://replays.springrts.com/player/152636/). This also affects how TS values are shown ingame. The exact value is shown only of players with privacy protection disabled.
0 x

raaar
Metal Factions Developer
Posts: 855
Joined: 20 Feb 2010, 12:17

Re: SLDB: Personal TrueSkill history graphs

Post by raaar »

Thanks.

This is interesting. I just skimmed through the descriptions. It takes into account predicted differences between players and the result (win or loss)

What the privacy setting actually does is weird, though.

It could show a "?" publicly or something. Instead it shows a rough estimate: That's why many players seem to have TS rating that's a multiple of 5.
0 x

dansan
Server Owner & Developer
Posts: 1198
Joined: 29 May 2010, 23:40

Re: SLDB: Personal TrueSkill history graphs

Post by dansan »

The "rounded" numbers are a historic relic that are used (or maybe not? ask bibim) to archive backwards compatibility with lobby clients.
They are the values that were previously assigned as values to the time-based skill that was originally used for balance calculation.

Code: Select all

1-bar == 20 TS    <-- maybe doesn't start with 20, but you get the idea
2-bar == 25 TS
3-bar == 27 TS
...
It can be found somewhere in the SPADS code.
So previously a x-bar was translated to y-TS and then the sum of team-TS (and TS-distribution) was used for balancing - just like now :)

That is where those funny numbers come from.
0 x

Post Reply

Return to “Lobby Clients & Server”