pls improve how balance is computed

pls improve how balance is computed

Classic game design, maintained to please you...

Moderator: Content Developer

Post Reply
Fuel
Posts: 114
Joined: 08 Jul 2006, 13:21

pls improve how balance is computed

Post by Fuel »

My impression is that currently the balance system sometimes makes little or no differentiation between these kind of players:

Rookie: really just started playing Spring BA, barely knows how to play, how units and resources work
Amateur: playing for a long time, builds units & eco, not so good at attacking enemy
Apprentice: Amateur + decent attack skills

Maybe not the best way to lay down the skills for this kind of players but you get the idea.
It looks to me that most of the players have TS 20 - TS 25 and very very few have TS higher than 25, and even much fewer have TS 27 or TS 30.
In other words, the system sometimes sees some Amateur or even Apprentice players like Rookies.
In team games, a TS 20 or TS 24 player which should be more like a TS 25 or TS 28, play an important role on the balance.
Even if they suck at attacking, they make units, they can delay an experienced enemy and with the help of an experienced ally can play an important role.

I am not saying all TS 20 players should suddenly be ranked as TS 30. What I am saying is I expected to see a more diversified range of TS values for players.

Maybe I am wrong, therefore can someone who has access maybe show the distribution of TS values ?
I am looking for a list of the percentages of players for each TS between 20 and 30, something like:
TS20: x1 % of active players (players which played in the last 3 months let's say)
TS 21: x2 %
TS 22: x3 %
..
TS 30: x30 %

I don't think this system is fair to people which like to improve their skills. If your TS gets too high compared to a big range of players, the balance for you starts to become a real gamble in the team games.

Thanks & I am looking forward for your thoughts
User avatar
Silentwings
Posts: 3720
Joined: 25 Oct 2008, 00:23

Re: pls improve how balance is comp

Post by Silentwings »

Trueskill values are dimensionless quantities and the relationship between pairs of them is (besides the usual Normal dist assumptions and Bayesian stuff) essentially defined by the fact that (1) TS cares only about win/draw/loss and (2) a teams skill is the linear sum of its players mu values. Changing (1) is not an option in Spring (balance systems should not be used to tell people how to play), so you are stuck with (2).

The "diversity" (by which I assume you mean deviation from mean of the set of players mu values) is not directly revelent to balancing; because the mu values are dimensionless they have no intrinsic scale. Although, because of the way TS tracks both mu and sigma values, they are indirectly related to the magnitude of change caused by a single update (controlled by an input parameter to the algorithm).

Changing the linearity assumption to alter the skewness (and suchlike, which would be relevent to balancing) would require a big change to the algorithm - probably a move away from Normal dist assumptions - and consequently a reasonably serious bit of mathematical work to calculate the new update rules. As i understand it, the Expectation Propogation approach would no longer be applicable and it would need a "proper" hidden Markov model.

To make a convincing case for changing it, afaics you'd need to design a new model with a new way of computing a teams TS from its constituent players, run data from a few hundred games to recompute skills (based on your new rules) and then be able to show that you can better predict game outcomes than before (for an example; look at what was done when TS was introduced). Obviously that would be a large piece of work. Since the TS system is widely used I would expect many people have already tried it and (since for us TS predicts very well and in general I've not heard much about extending it) mostly discovered that there was no major gain to make.

That said, I've no objection to the data being available, but since we have a privacy mode and most use it, the names would have to be removed (replace with #1,#2,...).
I don't think this system is fair to people which like to improve their skills. If your TS gets too high compared to a big range of players, the balance for you starts to become a real gamble in the team games.
I don't see anything in your post to support this claim. At present, the distribution of TS values is such that there is no difficulty finding ~matching linear sums of mu values to make teams out of. If games involving a small proportion of abnormally high/low skilled players seem more unpredictable, I would guess that its because the outcome becomes more determined on a single event i.e. whether or not player X lives or dies.

I'm not a statistician...
dansan
Server Owner & Developer
Posts: 1203
Joined: 29 May 2010, 23:40

Re: pls improve how balance is computed

Post by dansan »

I have created a statistic with gfx from the data on the replay site. Please be aware that bibims SLDB is the real database that knows smurfs and is more up2date etc. The TS data on a player on the replay site is only updated when he has played and a replay was uploaded. That said, the data should be mostly OK.

As the replay site has no clue about smurfs, I have created a "set" of the TS values. That removed all TS values that were equal to 2 numbers behind the comma. So if a player has a smurf whos TS changed more than that it will not be detected. That sounds like it wouldn't catch anyone, but it did eliminate half (!) the values... I still think there are quiet some smurfs in the higher numbers. If you want better numbers, ask bibim.

Image

0-15: 234
16-20: 431
21-25: 468
26-30: 371
31-36: 206
37-45: 114
Attachments
TS_distribution.png
(15.04 KiB) Downloaded 2 times
Fuel
Posts: 114
Joined: 08 Jul 2006, 13:21

Re: pls improve how balance is computed

Post by Fuel »

I'm not suggesting removing the current algorithm.
Maybe just manually override some values for some players and have algorithm start from there
Personally, the balance it's been a demoralizing factor for me which made me play much more less.
Fuel
Posts: 114
Joined: 08 Jul 2006, 13:21

Re: pls improve how balance is computed

Post by Fuel »

thanks dansan.
is this only for BA and for active users (played in the last 4 months lets say)?

I used to play Spring BA a lot in the last couple of years and I don't remember seeing aprox. 50 players with TS values around 35 value as the chart suggests.
so 2nd question is, does the data take into account several accounts of the same player?

edit: please don't include quotes of entire posts with large graphs.
Last edited by Fuel on 12 Jun 2014, 13:30, edited 1 time in total.
User avatar
Silentwings
Posts: 3720
Joined: 25 Oct 2008, 00:23

Re: pls improve how balance is computed

Post by Silentwings »

Maybe just manually override some values for some players and have algorithm start from there
Assuming that player did not change, they would simply converge quickly back to where they were (and they would do that because the algorithm is fair and effective...)
Kapytii
Posts: 64
Joined: 14 Mar 2013, 08:40

Re: pls improve how balance is computed

Post by Kapytii »

Balance is good when there are no new(real TS less than 10) players with TS showing 25(so what would that really be, 20-25?).

Why cant everyones TS be shows as exact? xx.xx, not the ~xx. Would make balance seem better I bet. That still would not help with the problem mentioned above :)

I know, not really what this threat was about.
dansan
Server Owner & Developer
Posts: 1203
Joined: 29 May 2010, 23:40

Re: pls improve how balance is computed

Post by dansan »

Oh sorry - didn't see this thread continued. My RSS-reader has this new strange thing, that spring-website only has max 5 entries... strange... and OT ^.^
Fuel wrote:is this only for BA and for active users (played in the last 4 months lets say)?
BA only.
Fuel wrote:I used to play Spring BA a lot in the last couple of years and I don't remember seeing aprox. 50 players with TS values around 35 value as the chart suggests.
There are at least 20 players with TS > 37 if you look at http://replays.springrts.com/hall_of_fame/BA/#Team
The HOF numbers are directly imported from SLDB and have no smurfs.
50 is probably to much, but 35 is a good guess.
Fuel wrote:so 2nd question is, does the data take into account several accounts of the same player?
dansan 12 Jun 2014, 13:18 wrote:Please be aware that bibims SLDB is the real database that knows smurfs and is more up2date etc. [..] As the replay site has no clue about smurfs, I have created a "set" of the TS values.
[..]
I still think there are quiet some smurfs in the higher numbers. If you want better numbers, ask bibim.
User avatar
Silentwings
Posts: 3720
Joined: 25 Oct 2008, 00:23

Re: pls improve how balance is computed

Post by Silentwings »

Why can't everyone's TS be shows as exact?
Because a significant number of people (myself included) prefer to use privacy mode and only show a rounded value.
User avatar
Jools
XTA Developer
Posts: 2816
Joined: 23 Feb 2009, 16:29

Re: pls improve how balance is computed

Post by Jools »

It doesnt matter how TS is shown. The exact value is still used for all calculations. Or I would assume.

Btw, that distribution doesn't really look normal. Maybe it's the Poisson distribution? Or the student's T-dist?
User avatar
Silentwings
Posts: 3720
Joined: 25 Oct 2008, 00:23

Re: pls improve how balance is computed

Post by Silentwings »

The exact value is still used for all calculations. Or I would assume.
Yes, it is.
Btw, that distribution doesn't really look normal.
Afaik there's no reason to expect the graph above to closely match any "common" distribution.
User avatar
Jools
XTA Developer
Posts: 2816
Joined: 23 Feb 2009, 16:29

Re: pls improve how balance is computed

Post by Jools »

Silentwings wrote: Afaik there's no reason to expect the graph above to closely match any "common" distribution.
You take a small sample of people whose TS-values follow the normal distribution (I think you defined that above) => you have the T-distribution.

If the TS-function isn't normally distributed then the µ and the σ of it can not be defined.
User avatar
Silentwings
Posts: 3720
Joined: 25 Oct 2008, 00:23

Re: pls improve how balance is computed

Post by Silentwings »

No. The values on the graph represent players (approximate) mean skill values, sampled from whatever distribution nature happens to have enforced on the general talent of Spring players. There is no reason to expect any special distribution for those, although anyone would expect a single humped distribution with reasonably short tails.

To be honest your post sounds like you are confusing the distribution associated to a single players observed skill (which is a modelling assumption of the algorithm, and is assumed to be normal) and the distribution of the mean skills of many players (which is what the graph above represents, and is chosen by nature).
User avatar
Jools
XTA Developer
Posts: 2816
Joined: 23 Feb 2009, 16:29

Re: pls improve how balance is computed

Post by Jools »

Maybe I am. But the observed skill is a sample from the population all player's skills, so if the population is normally distributed then a sample of it has the t-distribution.
User avatar
Silentwings
Posts: 3720
Joined: 25 Oct 2008, 00:23

Re: pls improve how balance is computed

Post by Silentwings »

But the observed skill is a sample from the population all player's skills, so if the population is normally distributed...
The graph posted above shows pretty clearly that, beyond superificial similarity, it isn't (and, as pointed out above, there is no reason why it would be). For example, there is a skew to the right and a flat-ish segment in the centre.
if the population is normally distributed then a sample of it has the t-distribution
That's wrong.

Likely you are confusing the way in which the t-distribution (with n-1 degrees of freedom) is involved in characterizing the sampling error of the sample mean of n independent samples from a standard normal. To be precise:
wikipedia wrote:If we take a sample of n observations from a normal distribution, then the t-distribution (with n-1 degrees of freedom) can be defined as the distribution of the location of the true mean, relative to the sample mean and divided by the sample standard deviation, after multiplying by the normalizing term \sqrt{n}.
Post Reply

Return to “Balanced Annihilation”