WN9 candidate prototype

[CraBeatOff 5,741]

3. The last time dumping the overall stat was suggested on WoTLabs, the idea was shouted down and the thread closed. Removing it entirely doesn't really solve anything though, except for persuading people that maybe recent stats are more valid. Recent stats acquired a bad reputation because they were so easy to pad with WN8: 1000 games should be good for WN9.

 

Consider me shouting it up. PR is more than enough for overall measures. If we dump development for over-all, we can ignore so much nonsense, and ultimately it doesn't matter right now how anyone played 1+ years ago. Leave PR and overalls to the people who have all the data and paid a consultant good money to implement it.  Focus on recents and recents only with a good formula. The only reason overalls were the basis for WN4-8 with Preator was because we wanted to get it into XVM to replace efficiencyv1 (which was innovative garbage) and XVM didn't have access to recent stats. XVM has switched to PR by default which is fine. Let them have PR/WN7/WN8/EFFv2 as settings and just focus on making something relevant and free of the noise of the past. 

 And also I'm not the only one -- crab as I recall got good win rates with crap dpg. 

Confirmed. All my solo efforts have yielded higher end of winning, with lower end of damage, except the BatChat25t 3 MoE run, which had both, but was executed with different intent. I probably could have had the same WR (67%) with 20% less dpg, but was trying to pad the damage. But I imagine that up until the abiilty to pad damage is its own skill, RN is right, and that it correlated well through the middle and upper, only breaking down at the extremes. 

[YuriPup 0]

I guess I don't play win8 well, win9 gives me a huge boost (over 10%). 

Is it the absurdly low plat rate?

W/R: 50.0%
Adj w/r: 48.6%
Plat ratio: 0.03
wn8: 995.2
wn9c1: 1144.2
wn9c2: 1124.6

[Necrophore 143]

I guess I don't play win8 well, win9 gives me a huge boost (over 10%). 

Is it the absurdly low plat rate?

W/R: 50.0%
Adj w/r: 48.6%
Plat ratio: 0.03
wn8: 995.2
wn9c1: 1144.2
wn9c2: 1124.6

It's not the platoon ratio. My WN9 went down with an even lower plat rate.

W/R: 54.9%
Adj w/r: 51.6%
Plat ratio: 0.00
wn8: 1523.0
wn9c1: 1403.7
wn9c2: 1323.3

(I really need to get the kids their own account someday. I think my recent is below my overall by now.)

[YuriPup 0]

Ouch...my 6 year old doesn't play (yet?) but he definitely get his own account...or use my wife's.

[GehakteMolen 2,063]

Buuuuuuut

Everyone is now purely looking at the platoon ratio, somethign which barely matters, since if i understand it correct, its only to get a rought estimation of solo vs platoon players, and use the dmg these solo players do for expected stats (roughly said)

So everyone saying its too high / low, its doesnt rly matter

ps: and 1 means heavy platooning, with 1 player, so if you play 50% solo and 50% in a triple platoon, you (should) get 1 (if not more), even if you play 75% of the time solo, you still very easy get ~0.5 platoon ratio, pure due to how heavy triple plutons weight... (@richard, pls correct me if im wrong )

[elFred 3]

I don't really figure how the platoon influence is integrated and used.

My platoon ratio is 0.88 is ,32 BIA for 22K games.

 

A "good" player I met last night (EU server obviously) : 

christopherx platoon ratio 0.82 , 244 BIA for 14K games (wtf less than me ??)

Now check his most played tanks and please explain me why he is rated "good"*

 

 

 

* and if someone can explain me why I cannot check my rating on wotlab , he will be welcome too

 

[RichardNixon 835]

Buuuuuuut

Everyone is now purely looking at the platoon ratio, somethign which barely matters, since if i understand it correct, its only to get a rought estimation of solo vs platoon players, and use the dmg these solo players do for expected stats (roughly said)

Actually the expected values don't use the platoon filter, partly because it's too inaccurate over low game counts. There's also an assumption that on average, players platoon a similar proportion of their games in each of their recent tanks, in which case the results would be largely unaffected.

If there was a systematic bias here, it would have a much stronger impact on tank-adjusted winrate than on tank-adjusted damage or WN9. Given that the winrate-relative errors don't spike at the high end, any platoon bias in the WN9 expected values is likely to be small.

 

The platoon filter was used to help derive the formula, although other considerations were arguably more important than the regressions. The choices in the formula are actually very limited.

The platoon filter was used to generate the scale & error graphs, as tank-adjusted winrate is useless without it. If I could improve the platoon ratio formula, then I could loosen the filter and extend the graphs more accurately into the superunicum range.

To put it another way, platoon ratio is simply an indicator (and not currently a very accurate one) of whether tank-adjusted winrate is useful for a player.

 

My platoon ratio is 0.88 is ,32 BIA for 22K games.

Your CCs account for most of your platoon ratio. Essentially the model thinks that you're about 25% double-platooned, 25% triple platooned and 50% solo. Based on known flaws in the model, both the BIA and CC components are likely to be overestimates.

[Garbad 10,517]

Why not 5k or 3k?  Or last 12 months.  That's enough to make padding a chore but still recent enough to reduce the legacy stats on most active players.

[breeeze 446]

You might wanna cap adjusted winrate, my Type 64 adj. WR is 111.2%, my T-50s is -18.2%

[RichardNixon 835]

Why not 5k or 3k?  Or last 12 months.  That's enough to make padding a chore but still recent enough to reduce the legacy stats on most active players.

I can certainly see why you'd want 3-5k for clan entry requirements, but it's down to what stat-sites are willing to store & display. Noobmeter has a half-account recent value, for example, and you can dig up shorter intervals manually.

It's possible that WoTLabs is already storing sufficient data for longer intervals, but it's not displayed. You'd need to ask Never.

 

You might wanna cap adjusted winrate, my Type 64 adj. WR is 111.2%, my T-50s is -18.2%

Nah, I should put an error value on it though.

[elFred 3]

<snip>

Your CCs account for most of your platoon ratio. Essentially the model thinks that you're about 25% double-platooned, 25% triple platooned and 50% solo. Based on known flaws in the model, both the BIA and CC components are likely to be overestimates.

I rapidly checked your code (very quickly) in the PlatoonRatio method you doesn't weight by battle count or I'm mistaken ?

 

IMO you need to change that

 

[Android25 962]

Maybe this is a terrible idea... I don't really feel like writing a script right now to see how it would affect the population, so I'm not sure. But have you ever considered slightly adjusting the expected values on a per player basis based on the percentage of games the player has in each individual tank?

I can think of two different ways it could be implemented. Either as a blanket up and down adjustment for all tanks. For example, the expected values for every tank are multiplied by X, where X is is a number between 0.95ish and 1.5 that scales between a tank being played for 0% of total battles, up to 100% of total battles. In this case, tanks with very few battles on the account would be treated with very slightly lower expected values (to account for learning the tank, upgrading it and the crew, ect.) and tanks with a high percentage of the overall battles on the account would be treated with higher expected values (to account for the fact that if you play a tank 30% of the time it should be expected that you perform better in it). With this method players would be rewarded more for playing a lot of tanks very well, than just a few.

or as an upward adjustment only for tanks that have been played more than "average" on the account. For example, if a player plays in one tank for a considerable percentage of their total games... say 15% maybe, the expected values are multiplied by X, where X is a number between 1 and 1.5 that scales between around 15% of battles and 100% of battles. Lets be honest, if a player plays only one tank in their account, it's to be expected that they would play better (in this example 1.5x good) in it than average, and the expected should reflect their use of the tank, not just the overall population.

 

I think that the first example would give a better scaled result, but I would love to see how either affects accounts that excessively pad tanks their very good in to artificially boost their WN8. Players like Garbad for example would be almost unaffected by the change, because his most played tank accounts for only 3% of his account, but (and this is just going from players off the top 10 list) players like Emperor (with 27% of his battles in the WTE) or Kancolle (with 42% of his battles in the 62A) would be punished with slightly higher values in those tanks played far more than average.

I don't think it should punish people for playing tanks they like a lot of the time... so I can definitely see the how it could have its faults... but there may be a better way of implementing it that would negate that that I simply haven't thought of.

 

Edit: I've given this a little more thought through the morning... You'd also have to set a multiplier per tier, because changing the expected for a tier 1 by 5% isn't nearly as dramatic as changing the expected for tier 10 by 5%. Because honestly the point would be to punish seal clubbing just as much as single tank padding...

Also, a base multiplier curve would look something like this where X is percent of battles the tank accounts for, and Y is multiplier. Notice that it meets 1 almost instantly. In this way, even accounts with 20k battles would only see the effects of lower expected values for the first 50 battles in a tank.

You could also set A (the first number 0.998) to 1, and there would be no lowered expected values.

[jacg123 656]

Wn9c2 is supposed to reduce the bonus received in wn8 for playing light tanks, but in the current implementation my RU-251 is my new highest tank @ 2813.1. In fact, half of my top 20 tanks (sorted by Wn9c2 and ignoring low sample size tanks) are light tanks. Is that intended or is there an error?

:edit: or am I just an outlier?

[RichardNixon 835]

Wn9c2 is supposed to reduce the bonus received in wn8 for playing light tanks, but in the current implementation my RU-251 is my new highest tank @ 2813.1. In fact, half of my top 20 tanks (sorted by Wn9c2 and ignoring low sample size tanks) are light tanks. Is that intended or is there an error?

:edit: or am I just an outlier?

I can't see any likely systematic error in that direction, other than the possibility that the NA metagame is sufficiently different for light tanks to get a small boost. You're talking about a pretty small numerical range here. It's also possible that you're slightly better at light tanks (according to WN9) than the average superuni.

 

Maybe this is a terrible idea... I don't really feel like writing a script right now to see how it would affect the population, so I'm not sure. But have you ever considered slightly adjusting the expected values on a per player basis based on the percentage of games the player has in each individual tank?

I actually do this to generate the expected values table, although it's based on XP earned rather than battle count. The general rules are that the effect is much larger for low tiers, artillery and non-premium tanks. Hence it appears to depend more on crew experience than player experience. IIRC this was true even when using battle count rather than XP, although in that case it's more difficult to factor out the equipment grinding.

The method is somewhat valid in a metric, but I'm not sure it's worth the complexity. It's a bit of a pain for recent values because you need the overall battles for each tank as input, not just the interval battles. It's not much of an issue for e-peen comparisons because having half your games in one tank is such an obvious profile flaw, especially now that stat sites are displaying per-tank values.

 

I rapidly checked your code (very quickly) in the PlatoonRatio method you doesn't weight by battle count or I'm mistaken ?

The final calculations don't have any battle weighting because the method is topgun-normalized. Having 200 topguns and 100 BIAs would give you the same ratio as 100 topguns and 50 BIAs. It's an attempt to adjust for potential errors in the model: If it overestimates both topguns and BIAs then it doesn't matter.

In the newer version, the calculations for expected BIAs per topgun and CCs per topgun are battle-weighted per tank.

 

[DramaQQ 2]

tell me the logic of this wn9... why do I lose that much? Because I play B-C25t very much or what?

Personal Rating	10455
WN8	2973.2
WN9 c1	2182.9
WN9 c2	1994.8

this rating is stupid. PR of >8800 is unicum level, I have 10455... almost 3000 wn8 is super unicum, wn7 is >2000 etc. but in wn9c2 I wouldnt even reach UNICUM stats haha :D. Seriously, just because I play BC25t and T71 etc.? that's crazy.

So I am everywhere at least unicum and super unicum, playing very skillful and what? I am blue with wn9c2? Jeah Sure

[GehakteMolen 2,063]

tell me the logic of this wn9... why do I lose that much? Because I play B-C25t very much or what?

Personal Rating	10455
WN8	2973.2
WN9 c1	2182.9
WN9 c2	1994.8

this rating is stupid. PR of >8800 is unicum level, I have 10455... almost 3000 wn8 is super unicum, wn7 is >2000 etc. but in wn9c2 I wouldnt even reach UNICUM stats haha :D. Seriously, just because I play BC25t and T71 etc.? that's crazy.

So I am everywhere at least unicum and super unicum, playing very skillful and what? I am blue with wn9c2? Jeah Sure

no, Wn9 uses scaling as Wn7 (more or less) 2000 wn9 = unicum

So if you would loose something, its because wn8 farming gets punished (like light tanks and some mediums)

[DramaQQ 2]

ah okay so its not worth to play mediums and lights... i have to play heavy tanks so arties who gets a bonus in wn9 can f*** me hard. Srsly if you are good in a medium tank and i am - dont think i am arrogant, but - i am a very good medium player and there are not many people who have 3300dmg/avg with BC, I get punished.. 

wn8 is simple: there are "avg." stats and if you play above avg. you get a very good rating.

Now I am a blue player!!! from Super Unicum to blue :D. It cant be

Actually I dont know why I am writing this here, i truely give a f***, because this game gets worse etc. Anyway... this "rating system" is kinda not good atm

[Siimcy 3,522]

Checked some guys and this rating seems to favor arty players pretty well. Time to play arty? Kappa

[Rexxie 10,788]

ah okay so its not worth to play mediums and lights...

 

Lights and meds still give great scores. Your T71, Cunningham, and T10 meds are basically the only things propping up your WN9. The problem is they aren't doing it as much as they did in WN8, and your other tanks look even worse now.

[Android25 962]

I actually do this to generate the expected values table, although it's based on XP earned rather than battle count. The general rules are that the effect is much larger for low tiers, artillery and non-premium tanks. Hence it appears to depend more on crew experience than player experience. IIRC this was true even when using battle count rather than XP, although in that case it's more difficult to factor out the equipment grinding.

The method is somewhat valid in a metric, but I'm not sure it's worth the complexity. It's a bit of a pain for recent values because you need the overall battles for each tank as input, not just the interval battles. It's not much of an issue for e-peen comparisons because having half your games in one tank is such an obvious profile flaw, especially now that stat sites are displaying per-tank values.

Aren't recent values just going to be calculated from overall delta's? Unless there's a recent data API function that works for everybody that I don't know about, stat sites are still going to have to pull the overall data and compare it to the last saved overall to get the delta's, just like recent is currently calculated. Total functions wise, it would be 6*(number of tanks)+1, which, computational time wise, nobody would notice.

I understand that you can just go to a players account and immediately say "well this shouldn't count...". I just feel like if it's that easy to eye it, there should be something that can be mathematically done to keep said players stat low in the first place.

There must be a somewhat reasonable way to weed out seal clubbing, rerolling (I guess it doesn't really matter), and simple mass padding...

[RichardNixon 835]

There must be a somewhat reasonable way to weed out seal clubbing, rerolling (I guess it doesn't really matter), and simple mass padding...

WG are already beating up on tier 1-3 sealclubbers, and the difference between a 3 and 5 skill crew is small, so mass-padding a single tank doesn't have a clear advantage except in non-premium mid-tiers, where you're generally talking about a ~10% advantage. The mid tiers are generally rated higher than in WN8 due to the grind adjustment, so it's tougher than it was.

There's a secondary argument that players who are (genuinely) good at one tank should be ranked lower than players who are good at a lot of tanks, but that may be getting a little too ideological.

[fabriciohss 7]

WN8	2514.1
WN9 c1	1969.9
WN9 c2	1859.4

Ouch. Thats a huge drop. 

To be fair Im ok with it.

 

 

[jacg123 656]

ah okay so its not worth to play mediums and lights... i have to play heavy tanks so arties who gets a bonus in wn9 can f*** me hard. Srsly if you are good in a medium tank and i am - dont think i am arrogant, but - i am a very good medium player and there are not many people who have 3300dmg/avg with BC, I get punished.. 

wn8 is simple: there are "avg." stats and if you play above avg. you get a very good rating.

Now I am a blue player!!! from Super Unicum to blue :D. It cant be

Actually I dont know why I am writing this here, i truely give a f***, because this game gets worse etc. Anyway... this "rating system" is kinda not good atm

Let me get this straight... You made an account to complain about how the new *potential* rating system caught you red handed with a ton of tanks that have bad stats?

Also, pls stop clubbing seals with your cunningham.

[dirtd0g 1,181]

WN8	2514.1
WN9 c1	1969.9
WN9 c2	1859.4

Ouch. Thats a huge drop. 

To be fair Im ok with it.

 

 

Not really.

If I'm understanding this correctly, a WN9x of ~2000 is unicum...

[CraBeatOff 5,741]

Not really.

If I'm understanding this correctly, a WN9x of ~2000 is unicum...

Alex, what is scaling? 

That's correct dirtdog. Pick the next category.

I'll take Reading The Fucking Thread for 800.