#Starting TS should be lower

1 messages · Page 1 of 1 (latest)

cobalt folio
#

Starting TS is currently 17 or so. This is way too high and not representative of the skill level of new players. Something like 10 would be better. With the current starting value it's possible to get "balanced" teams like this:

image.png
placid brook
#

That would just skew all OS values lower, its been talked about many many many times. If you want starting OS to update faster, dont play large games. The system just looks at win/loss so its hard to determine who is doing what inside the games. The more players introduced, the less it is certain that a particular player made a difference toward that win. If a new player played a few 2v2s, their OS would adjust accordingly in just a few games.

placid brook
#

if you look at where they experimented to get their results, they say:

#

which is basically what we do I believe

#

in other words, it wont matter

#

although there is some merit to finding the actual average of all players and changing initial value to that since they say its supposed to be the mean

#

but honestly that is probably the average anyways

#

or close to enough that it doesnt matter

dense compass
#

its begun again

dense compass
tardy aspen
cobalt folio
#

Increase uncertainty to 15 then?

cobalt folio
#

Is uncertainty decrease weighted as well? Maybe 1/8th for 8v8, 1/7th for 7v7 etc.

coral light
#

The issue is the open skill model + assumptions. As has been said, if you introduce lower-skill levels for new people into the open skill model, you're just going to shift the entire distribution downwards. I have just started researching/developing alternative skill rating metrics, but that is a long term personal project and not officially affiliated with BAR (I'm using BAR data and obviously would like to give back to the community as much as I can).

dense compass
grim gale
#

man remember playing with new players back before the OS days

#

Like.... 9 months ago or whatever lol

#

now THAT was bad

dense compass
#

that balance algo didnt take into account uncertainty

radiant mauve
dense compass
radiant mauve
#

Isn't the server doing team balancing by using a match quality test function?

#

Because a match quality function takes sigma and mu into account, but it isn't as simple as saying (mu -sigma).

#

The effect of sigma isn't linear, which isn't easy to represent as a single number.

#

Unless the seeding is using that value, which may make sense.

#

Well if it works...

placid brook
#

Yall always against kicking noobs for some reason but love to complain when they on your team

cobalt folio
#

The problem is not the noobs being in the game. It's that one team has 5 noobs and one team has 1 noob.

#

The teams from the screenshot I posted didn't actually fire, because someone 'dodged'. But the fact that the balancing algo can come up with that is a bit silly.

placid brook
#

You receive less of a rating change on teams where you're expected to lose

#

Also that's the reality of mixed skill lobbies

#

Balance algorithm does not consider game time

grim gale
#

I think that with the rate of growth so high relative to the size of the established player base, there's just not a whole lot that can be done to balance with new players without sacrificing rating quality for everyone

thorn marlin
#

Combine ingame, spec time which is basically the ranking system with OS and uncertainty to calculate a new rating value

placid brook
thorn marlin
#

Ingame time could be a linear graph crossing zero for 100 hours. If you’re under that amount played you’re considered less experienced and if you’re above you’re added xp

#

Add or deduct accordingly from OS and uncertainty

compact stratus
#

i think 17 is more than enough

#

already brought it down

#

no need for more

thorn marlin
#

You can’t just give new players 17. what about the noobs from last week? They’ll still have the higher value

ocean cipher
#

From what I have seen new player os adjusts fairly quickly to a sane value, at least if they play some non 8v8 games

#

We are about to introduce skill restricted rooms soon so it will, we can also think about restricting big room to players with at least 5h in-game time as we did on official lobby server I guess

#

They could be played hours or spec hours. Just so that the total newbies don't cause unbbalance in these rooms before their os adjusts to a decent value

#

Would also mitigate some of the smurf issues

cobalt folio
#

Or just make players win 1v1 vs Brutal AI

ocean cipher
compact stratus
plain saffron
#

I've zero plans to lower starting TS

placid brook
#

(25/3)^2 is what the paper says, comes out to 17.33
seems fine to me tbh

thorn marlin
#

I thought it was 25

plain saffron
#

No, that's starting skill. Rating is a different value

#

OpenSkill algorithm uses a few terms such as mu and sigma which can be confusing for those unfamiliar. As such BAR interface uses different names, this post is a reference for the different terms.

mu or skill - One of the two values stored to represent player skill. Winning games will typically increase this value.
sigma or uncertainty - The second of two values and is used to represent player uncertainty, the higher the number the less certain the algorithm is about the player. Playing more games will lower this value.

game rating - The value used for game balance, it is defined currently as skill - uncertainty. This means given the defaults of skill = 25, uncertainty = 8.3 new players have a game rating of 16.7
leaderboard rating - The value shown on leaderboads. Defined as skill - 3 * uncertainty meaning everybody starts with a leaderboard rating of 0.

compact stratus
#

been 2-3 months since its been to 17

#

and i dont want another reset when current balance is fine

#

and another lowering of starting os will just bring more imbalance as then we have to reset the skill like we should have when we changed to 17

thorn marlin
#

If 25 is skill we can still add ingame time to the equation

plain saffron
#

I do not intend to factor ingame time into any calculations regarding player ratings.

#

While there might be a correlation between time and skill, it's not a causal relationship and it would be a mistake to treat it as such

vernal dock
#

Some games get around this by forcing new players to do some "placement games" before allowed in other ranked games. E.g. 5 games 1v1 against opponents of varying skill so the ranking can adjust. Unfortunately not possible for us with current playerbase obviously.

plain saffron
#

I have plans to do something like that with matchmaking, for public games though we've no intent to put blockers in place as a common play method is someone bringing their friends on and wanting to play with them.

placid brook
vernal dock
#

I think the idea in most rating systems is that any rating gained has to come from another player. So you cannot reward rating for winning against bots. Otherwise someone could just create a bunch of 40 OS accounts out of nowhere and ruin the rating system with those.

placid brook
#

fortnite has this system tho, there arent always lowskill/new players playing to fill 100 person lobbies all the time so they populate your first few games with bots so you get wins and once it determines your skill you get put in with a bunch of players

plain saffron
thorn marlin
#

In BA low level players can’t chose their start position. They get the spots that are left

ocean cipher
#

And it's only for the casual big team room/s basically where managing these new players become an annoyance and disrupts the balance for many at a time

plain saffron
#

I'm against new players not being allowed to play in public games just because they're new

dense compass
#

the players in the room are the ones that should be voting if they want the new player or not IMO
that is one things players in lobbies can do that gets fairly well ignored

plain saffron
#

If you don't want new players playing in your game then use the commands for setting levels. Kicking people because they are new could be seen as bullying.

dense compass
#

yes that is the cost of having forcespecing people with others in the lobby
its more that people dont realise that they have this as an option but come complain about new players ruining balance instead
new players need a means of becoming not new
dont get me wrong this means of player control isnt the best either

plain saffron
#

I've just updated the server to now show play bounds (if set) to anybody when they join a lobby

thorn marlin
#

I just witnessed an 8v8 game where one player didn’t even know he had to build mexes

plain saffron
thorn marlin
#

I never kicked a noob. I gave instructions on what to do

compact stratus
#

i suggested 1 thing a year ago

#

make tuturiols

#

and make tutorials compulsory to enter multiplayer

cobalt folio
#

Personally I don't have a problem playing with noobs. But this 'balance' was obtained using the currently algorithm and anyone can see it's clearly unfair. What can we do to fix the balancing algo?

image.png
compact stratus
#

so the guy isnt a total noob

plain saffron
compact stratus
#

does my thing makes sense teif

#

or against that too?

plain saffron
#

I'm not against it in principle but I'd need to see the tutorials and think about it further before saying yes.

compact stratus
#

thats fine

#

im only suggesting this cause every other major rts already does it

#

and it works well

plain saffron
#

I fully agree a tutorial being completed before online play is allowed as a requirement can be a good thing, just depends on the tutorial and context

thorn marlin
#

Minimum experience rooms will be enough. If this stays a problem I’ll just create a room where noobs can’t play

#

If someone is new let them ply with their friends or do the academy. Don’t shut down mp in general for them

ocean cipher
compact stratus
#

it is the one that works tho

#

think about it

#

people will learn how to play the game before entering multiplayer

#

then people inmultiplayer dont need to teach them from scratch just so they get balanced games

ocean cipher
#

@compact stratus take a listen

brisk aspen
plain saffron
#

There is no option for that, just game rating. I am not adding an option to exclude people just because they are new, that's incredibly toxic.

#

Also, who has 27 game rating after winning one game?

placid brook
#

Don't you need boss to set minimum room level? Seems excessive no? Why not just have votes

plain saffron
#

Because server commands don't have a vote interface

patent flicker
#

They could be played hours or spec hours.
Are you serious?

dense compass
#

its already possible to do with boss mode

patent flicker
#

That is not why I asked...

cobalt folio
# plain saffron That screenshot is useless for debugging, next time please chat `$explain` and c...

[19:34] Coordinator: ---------------------------
[19:34] Coordinator: Balance logs, mode: solo
[19:34] Coordinator: Tried grouped mode, got a deviation of 6 and reverted to solo mode
[19:34] Coordinator: Picked StarDoM for team 1, adding 38.47 points for new total of 38.47
[19:34] Coordinator: Picked noidzz for team 2, adding 21.78 points for new total of 21.78
[19:34] Coordinator: Picked Bugious for team 2, adding 21.32 points for new total of 43.1
[19:34] Coordinator: Picked ta_comm for team 1, adding 20.6 points for new total of 59.07
[19:34] Coordinator: Picked ClaudeLeLezard for team 2, adding 18.05 points for new total of 61.15
[19:34] Coordinator: Picked roh_cirtep for team 1, adding 17.11 points for new total of 76.19
[19:34] Coordinator: Picked dldbstn9494 for team 2, adding 16.91 points for new total of 78.06
[19:34] Coordinator: Picked TArules for team 1, adding 16.81 points for new total of 92.99
[19:34] Coordinator: Picked andorDrakon for team 2, adding 16.76 points for new total of 94.82
[19:34] Coordinator: Picked SYNTHETIC for team 1, adding 16.6 points for new total of 109.59
[19:34] Coordinator: Picked Kylaxezu for team 2, adding 14.98 points for new total of 109.8
[19:34] Coordinator: Picked Taming_Strange for team 1, adding 13.71 points for new total of 123.3
[19:34] Coordinator: Picked tremolol for team 2, adding 10.98 points for new total of 120.77
[19:34] Coordinator: Picked ProAustralian for team 2, adding 10.85 points for new total of 131.63
[19:34] Coordinator: Picked Wookieguy for team 1, adding 10.48 points for new total of 133.78
[19:34] Coordinator: Picked BilboTBaggins for team 1, adding 7.36 points for new total of 141.14
[19:34] Coordinator: Deviation of: 7
[19:34] Coordinator: Team 1 - sum: 141.1, mean: 17.6, stdev: 8.78
[19:34] Coordinator: Team 2 - sum: 131.6, mean: 16.5, stdev: 3.85
[19:34] Coordinator: ---------------------------

image.png
#

Team 1: R7, R4, R4, R4, R4, R2, R2, R1
Team 2: R4, R4, R2, R3, R1, R1, R2, R1

placid brook
#

super mixed skill, you have extremely good players like stardom and completely newbie guys what do you expect

dense compass
#

skill ranking system are always reactive to what they have
they can never prescibe a rank to a player
even if there are placement matchs
the rankings are always reactive to data

dense compass
cobalt folio
#

StarDoM ended up the team with only 2 R4s and lost

image.png
dense compass
#

stardom represents roughly + 20 points over the next best player

cobalt folio
#

Yes, and yet in the original balance he still got the better other 6 players

coral light
#

in klu's defence, on a rank aspect team 1 has a literal absolute advantage over team 2

#

I.e. for team 1, every member is >= to the opposing member of team 2

#

(in terms of rank)

cobalt folio
#

Delete top player and you get:

image.png
#

Except top player on the better team (left) is StarDoM.

dense compass
#

this is more of a scheduling thing for the algo to deal with

plain saffron
# cobalt folio [19:34] Coordinator: --------------------------- [19:34] Coordinator: Balance lo...

Thank you for sharing. I can't see an issue here, there are new players in the game and they're being rated accordingly.

For those saying new players should be playing only with new players, when they leave that pool they will have either inflated or deflated ratings and ruin games even more. Two terrible players always playing against each other where one wins 80% of the time, that one will have a very high rating but still be a terrible player.

oblique lintel
#

This is the case in most games btw - it's just that in other games your rank is hidden for 5 or 10 games and you're instead "uncalibrated" or in "placement" phase

#

you still have a rating, and you will get put against players of vastly different skills during those games (especially the first few)

plain saffron
#

Now that is an interesting idea

oblique lintel
#

other games do a few little tricks like this to avoid too many players getting mad, on top of hiding their ranking algorithm, hiding their rank icon mmr thresholds, not showing players their precise rating to the decimal value. many games work exactly the same as ours (or at least close to) - it's just they hide stuff from you and we lay everything bare for you to scrutinise and get angry about