bjshnog

⟪WN8⟫ Development / Resources

2,170 posts in this topic

Generally, this ratioWins value is broken. We'll work out some other way to calculate it, but it's actually rather difficult in this case.

 

It's in the previous post.

 

Sum(f(x), rangeMin, rangeMax, ind) = // I am aware this is abuse of notation. This is just for demonstration.

{var sum;

sum = 0

for(ind=rangeMin; ind<=rangeMax; ind++)

    {sum += f(ind)}

return sum}

 

Gotcha. Seems straightforward enough. Since you will be providing the tank's topdamage, topfrags, etc., this can theoretically be implemented into worldoftanks.com profile page since it provides the number of battles for each tank.

Share this post


Link to post
Share on other sites

Because server top winrates at each tier are extremely variable, we need to come up with another solution. I think WR is best done if split up into tiers. After this, we can give each tier a weight (which could just be the tier number, but I think a curve would be better) and a comparison value, for example:

  • Tier    I:  1, 93%
  • Tier   II:  2, 84%
  • Tier  III:  3, 76%
  • Tier   IV:  4, 69%
  • Tier    V:  5, 63%
  • Tier   VI:  6, 58%
  • Tier  VII:  7, 54%
  • Tier VIII:  8, 51%
  • Tier   IX:  9, 49%
  • Tier    X: 10, 48%

This curve could be easily changed based on WR stats of a bunch of average high tier players who play a lot of low tiers too (not me though, because I haven't played tier 1 for ages and I still have low win rates in those tanks). The reason why it should be for average players is because for the server top, the difference in WR values at each tier will be much less than it would be for an actual average player, though we could possibly use the top stats. In fact, it may be better to just come up with artificial values and see how the curve fits with some players like Marxist, _PanzerVor_, and other seal-clubbers, who have large tier ranges.

 

After this, we would take the total wins at each tier, find the ratio of WR per tier to comparison WR for that tier, multiply it by number of battles, then by tier weight, add up the total at every tier, then divide by the total battle weight (sum of battles at each tier multiplied by their respective weights), to get the final weighted win rate.

 

Finally, we use a function to transform it into a curve which has wWR [0,1], then multiply it by weightWins, for the final WR value for WN8.

Share this post


Link to post
Share on other sites

Sounds and looks good. Now we just need the data. XD

Share this post


Link to post
Share on other sites

Playing T49 and looking at it´s average stats has made me think that it´s complicated to use a single expected value for each tank for each stat. This is due to skill floor/cielings explained by Crab in his post... which is somewhere I can´t find...

 

 

Anyways, we will have to figure out what is the best value to normalize by, but I am thinking an average of (floor+cieling)/2 would be best. Floor could be avg stats and cieling could be top 100, top 1% , top 5% or something similar.

Share this post


Link to post
Share on other sites

I was thinking of doing that. I'm pretty sure I posted something about it in this thread...

 

EDIT: Nope, I thought about it, but didn't post it.

Share this post


Link to post
Share on other sites

... It is also because some vehicles in classes have very different scouting capabilities. For example, the Loewe has a high view range and tends to insta-spot the enemy team on the Standard Malinovka map, while the Tiger II doesn't. Maybe that was just my crew and modules, but in every battle, if I saw an enemy Loewe or Pershing, I turned right around and got the hell out of there or died.

 

 

Löwe and KT have the same base viewrange - 400m. It just depends on modules, every tank with 400m + Optics can spot above the field in campinovka.

Share this post


Link to post
Share on other sites

Löwe and KT have the same base viewrange - 400m. It just depends on modules, every tank with 400m + Optics can spot above the field in campinovka.

 

I see. Well, that's what happens when I don't check my stats. :3 I have Coated Optics on my Löwe.

Share this post


Link to post
Share on other sites

Playing T49 and looking at it´s average stats has made me think that it´s complicated to use a single expected value for each tank for each stat. This is due to skill floor/cielings explained by Crab in his post... which is somewhere I can´t find...

 

 

Anyways, we will have to figure out what is the best value to normalize by, but I am thinking an average of (floor+cieling)/2 would be best. Floor could be avg stats and cieling could be top 100, top 1% , top 5% or something similar.

 

I was thinking about this for a while and remembered something. Tier 1 averages are much, much lower than top stats compared to at tier 10. If we just take the average, then unicum sealclubbers would get a massive bonus just for playing at tier 1. Of course, we could just add a term which is a function of tier to each expected damage, or a multiplier. Then again, the stats remain to be received. If I could read the post you mentioned, that might help me understand this more, but at the present moment, it seems to be flawed.

 

EDIT: I found the

 

While the idea makes sense in theory, it won't work as a raw average of floor and ceiling, due to the vastly different ratios of floor and ceiling at lower tiers. We would be better off with (stat - floor) / (ceiling - floor), and we would probably have to use a value other than the average stat for the floor. Maybe the average stat divided by 2.

Share this post


Link to post
Share on other sites

Hm... let me think about this, but you are right, avg+top/2 will not work...

Share this post


Link to post
Share on other sites

(stat - floor) / (ceiling - floor) won't work either. The more I think about it, the more it seems like simply using the ceiling value is the way to go. Then again, the skill floors are still there...

The problem is that average players won't necessarily have average stats at low tiers, and if they decide to play lots of low tiers later on, then their stats in low tiers will show that they are way above average.

Share this post


Link to post
Share on other sites

Yeah, you are right. We just have to do some analysis once we get the data to choose how the top will be calculated in order to make it as accurate as possible. Maybe try with different % of the population as tops, and check how each of them work.

Share this post


Link to post
Share on other sites

I was also thinking. Even if someone overperforms top100 in a tier 1, should that be worth the same WN8 as in a tier 10?

 

Let me rephrase that. We all know playing tiers 1-6 can get you huge stats compared to high tiers. Now, if we implement this "sort-of-per-tank" weighted rating for WN8, differences between tanks will dissapear since your performance on each tank will be (kinda) measured against how well the rest of the server performs on it.

 

Lets say the top100 for MS1 is 3 kills and 500 damage. Lets say top100 for Foch155 is 2 kills and 4000 damage.

Should a player who only played MS1 getting 4.5 kills and 750 damage get the same WN8 as someone who only played Foch155 and averages 3 kills and 6000 damage?

 

If not, how much less would he deserve?

Share this post


Link to post
Share on other sites

I was also thinking. Even if someone overperforms top100 in a tier 1, should that be worth the same WN8 as in a tier 10?

 

Let me rephrase that. We all know playing tiers 1-6 can get you huge stats compared to high tiers. Now, if we implement this "sort-of-per-tank" weighted rating for WN8, differences between tanks will dissapear since your performance on each tank will be (kinda) measured against how well the rest of the server performs on it.

 

Lets say the top100 for MS1 is 3 kills and 500 damage. Lets say top100 for Foch155 is 2 kills and 4000 damage.

Should a player who only played MS1 getting 4.5 kills and 750 damage get the same WN8 as someone who only played Foch155 and averages 3 kills and 6000 damage?

 

If not, how much less would he deserve?

 

I believe I already posted this:

 

A function to should be added or multiplied to all expected top values, such as [100+(-tier*2 + 20)]/100 * topStat, so that tier 1s require 18% higher stats for the same WN8 result, tier 5s require 10% extra, and so on. I noticed that sela has higher above top stats in his tier 1s than in his tier 10s, so I thought of this.

 

About tier, we can just scale the top100 value up by 2% per tier difference from tier 10. For example, for tier 1, the top100 values would be increased by 18%, 10% at tier 5, 4% at tier 8, 2% at tier 9, etc.

 

Found it.

Share this post


Link to post
Share on other sites

Yeah, now that you mentioned it, I remembered... that sounds about right...

Share this post


Link to post
Share on other sites

Some player compile stats in OP tanks after they're released, while others play them post nerf

 

Zakaladas and his VK3601 is a good example. Post release, pre-nerf konish gold spam (when the rounds were gold only), compared to someone who picks it up now? Same thing with the Lor 40 when it was tier 8, anyone who plays it now is at a disadvantage statistically. many other examples as well. I have read this entire thread and nothing much can be done? 

 

Also, thanks for the effort guys I loves the numbers

 

 

 

Share this post


Link to post
Share on other sites

Some player compile stats in OP tanks after they're released, while others play them post nerf

 

I have read this entire thread and nothing much can be done?

 

Pretty much. WG's API doesn't have historical data (ie. when different tanks were played), so we can only go by what we have.

Share this post


Link to post
Share on other sites

Talking about floors and ceilings, what if you set WN8 800 to be average values for your combination of tanks, 2000/2200/whatever to be the top values in those tanks, WN8 below 800 is basically proportional to the average, 800-2200 proportional to (ceiling-floor), 2200+ proportional to ceiling?

 

After all, when evaluating players near average, average power of a tank is relevant.  When evaluating purples, the ceiling is relevant.

Share this post


Link to post
Share on other sites

Talking about floors and ceilings, what if you set WN8 800 to be average values for your combination of tanks, 2000/2200/whatever to be the top values in those tanks, WN8 below 800 is basically proportional to the average, 800-2200 proportional to (ceiling-floor), 2200+ proportional to ceiling?

 

After all, when evaluating players near average, average power of a tank is relevant.  When evaluating purples, the ceiling is relevant.

 

It's not that easy. As I have already said, it won't work at all. Average players who play lower tiers will pretty much automatically get higher WN8 scores simply because the average damage at those tiers is so much more distant from any possible ceiling.

 

Right now, all we can do is make the preliminary formula and multiply it by a number to set the average WN8.

Share this post


Link to post
Share on other sites

I have the data from Phalynx. I need time to analyze this, for some reason it takes forever to calculate average values of damage and such per tank... wussy Excel can´t handle 750000 lines and 12 columns... 

 

And this is after I filtered the data to players with at least 50 games on tank.

 

Some previews:

 

 

 

And some problems:

Many tanks have a very low number of players, like these:

 

sjht.jpg
 
To have an approximate performance rating I used WN7, and checked what class was getting whacked the most. To my surprise SPGs are pretty balanced, even with the tier change.
 
ub0k.jpg
 
Lastly, WN by tier. Man those tier 4 tanks are bad. A-20 specially is dragging down the WN for it´s tier lol. Seems everyone sui-scouts in it.
 
jsgc.jpg
 
 
What I did was for each tank, take the stats for the top 10% players, remove top 1 %, and take the average for each stat. It seems pretty reasonable for most tanks, even for some tanks with low number of players. However, specially overperforming and underperforming tanks will be a problem. For extremely underperforming tanks, like the A-20, you can average 0.5 kills and 170 damage and you are already in the top 10% for the tank... o_o
Although the numbers dont seem that unreasonable, for the Lee an up. Maybe we could multiply all tank values by tier average or something to make them less extreme?
 
 
Low performers:
xubx.jpg
 
Top performers seems to be reasonable too, maybe Hotchkiss is a bit extreme also... 3.2 frags? Wow...
 
2qrg.jpg
 
 
 
 

Share this post


Link to post
Share on other sites

 

 

 

 

And some problems:

Many tanks have a very low number of players, like these:

 

 

 

 

what server are these numbers from?

 

 

edit:

 

and regarding arti, there you are right and wrong. For the avg tomato WN might be right but not for the server top. Here is a screen from my second account (edit ~97% solo random):

 

unbenannt2ssl9.jpg

 

 

after you sort out WN8 especially the damage part, there really should be something done to WN for the good arti players (top ~5%) because they really get punished.

 

 

edit 2:

 

I m now in possession of the dossier file from Baine [EU], many consider him the best arti player when it comes to the russian SPG-line on the EU-Server:

 

unbenannt0zuvj.jpg

 

I cant comment on % of random solo vs platoon/cw/tc,  but 1986 WN7 on 261 for the "best" arti player and even the Obj 212 has a "pretty low" number.

 

If Praetor77 or someone else involved with WN-rating wants to look at the dossier please pm me.

Share this post


Link to post
Share on other sites

Please upload both (and any other dossier you can get your hands on) to vbaddict.net. WE need a bigger database to be sure our numbers are not biased.

 

Arty will be balanced out pretty nicely by using tank specific stats. They usually get lower kills/game but higher DMG/game than same tier tanks, which in WN7 hurt them, but this should no longer be the case for WN8.

Share this post


Link to post
Share on other sites

Unfortunately, I can't see any of those images in your post, Praetor. Could you switch to dropbox from now on in this thread?

Share this post


Link to post
Share on other sites

As a comparison, I sorted by alphaWN8 and threw out top 1% and bottom 90%.

JAoQ9tY.jpg

 

Thank you for the delicious data!

Share this post


Link to post
Share on other sites

  • Recently Browsing   0 members

    No registered users viewing this page.