bjshnog

⟪WN8⟫ Development / Resources

2,170 posts in this topic

Sounds good. We would need a decent (aprox. 50k) player DB though, which we don´t have. If we can get HIbachi to modify wotstatgrabber and give us the data we need, we could use it to put together a DB.

 

So... seems like WN8's release is slowly closing in...

 

Anyway, my solution for tying stats together will also rely on correlations between stats. The result of it will be that if two stats are only weakly correlated with each other, and the player ignores one to focus on the other because it is more strongly correlated to wins (ie. ditching Defense to focus on Frags), then they will get some points, but they will also receive a very small penalty for ignoring that stat.

 

So, if Defense is worth 0.2 and Frags is worth 0.5, and the player has 1 of each, then their score would be 1*0.2+1*0.5 = 0.7 (because both stats are equal). Let's say Defense and Frags are correlated with a coefficient of 0.25. If they ignored Defense and went for Frags, resulting in 0.8 and 1.2, then they would get something like ((1+((0.8/1.2)-1)*0.25)^(1/3))*(0.8*0.2+1.2*0.5) = 0.738, whereas without the adjustment, it would be 0.8*0.2+1.2*0.5 = 0.76 (notice that it is only a small difference, because there is only a 0.25 correlation between the stats). If it was 0.5 and 1.5, then the result would be ((1+(0.5/1.5-1)*0.25)^(1/3))*(0.5*0.2+1.5*0.5) = 0.8, where without the adjustment, it would be 0.5*0.2+1.5*0.5 = 0.85.

 

This can be written shorthand as the following:

r = 0.25, 0.2:0.5.

1:1 results in 0.7 with or without adjustment.

0.8:1.2 results in 0.76 without adjustment.

0.8:1.2 results in 0.738 with adjustment.

0.5:1.5 results in 0.85 without adjustment.

0.5:1.5 results in 0.8 with adjustment.

 

A different situation:

r = 0.5, 0.2:0.5.

1:1 results in 0.7 with or without adjustment.

0.8:1.2 results in 0.76 without adjustment.

0.8:1.2 results in 0.715 with adjustment.

0.5:1.5 results in 0.85 without adjustment.

0.5:1.5 results in 0.743 with adjustment.

 

Basically, if two stats are correlated more strongly, then one would be expected to increase a little more with the other. If that isn't happening, then the player must be doing something wrong. In the case that a player has 1 and 1, and they increase that to 1 and 1.5, they will still get points for the extra 0.5 for Frags, but not as much as if they were to increase Defense also. With a stronger correlation, they would get slightly less points.

Share this post


Link to post
Share on other sites

To be honest, I think we should reanalyze cap points to see if they should be added. Logically, if you get any cap points in a battle, you win, so...

Share this post


Link to post
Share on other sites

Forget about these fucking cap points

 

More games are lost because of cap idiots, than by the "kill all" faction by far. Furthermore how do you want to seperate usefull cap points from the huge nonsenese farmed at the end of the match?

Share this post


Link to post
Share on other sites

Forget about these fucking cap points

 

More games are lost because of cap idiots, than by the "kill all" faction by far. Furthermore how do you want to seperate usefull cap points from the huge nonsenese farmed at the end of the match?

 

Good players who cap win more than good players who don't cap, and bad players who cap win more than bad players who don't cap. Also, if a player does nothing but cap, then their WN8 score will still take a massive hit. See three posts above.

 

Also, capping turns a noticeable amount of battles.

 

Even if it is only worth about 10% of damage, it will still make it more accurate.

 

And I'm sure everyone's seen those retards who beg to stop capping so that they can kill the last few tanks when they are at base, also capping.

Share this post


Link to post
Share on other sites

Good players who cap win more than good players who don't cap, and bad players who cap win more than bad players who don't cap. Also, if a player does nothing but cap, then their WN8 score will still take a massive hit. See three posts above.

 

Good player who cap more dont win more often, the just get a better Eff-Rating. Bad player who cap more often, lose even more matches.

Also, capping turns a noticeable amount of battles.

to a certain loss, I know. You can watch it every day on Ensk.

Share this post


Link to post
Share on other sites

Although i don't like the idea of cap points being added on a personal interest (i never cap so this won't do me any good ;)). I do actually see that capping isn't the best direct strategy for winning a game. But using it for other needs its actually really helpful. Forcing enemy's out of hull down positions to decap etc. the problem with this is that these cap points get decapped a lot so won't show in the results. The cap points that do show are hardly ever truly match making caps. More often than not these are the "its 14 against 5 but the red player still wants to cap" or the "oh noes the battle is almost over lets farm some cappoints"

Share this post


Link to post
Share on other sites

...But using it for other needs its actually really helpful. Forcing enemy's out of hull down positions to decap etc. the problem with this is that these cap points get decapped a lot so won't show in the results. The cap points that do show are hardly ever truly match making caps. ...

This!

@bjshnog:

I have had enough matches with nearly 300 decap points for myself against superior enemies forces and won. The myth that cap wins matches comes from people who either dont understand the game or Eff-Junkies. There are matches that can only be won by cap - no discussion there - but these are a small number, and often enough matches are lost by people who dont know when to cap and when not.

Share this post


Link to post
Share on other sites

To be honest, I think we should reanalyze cap points to see if they should be added. Logically, if you get any cap points in a battle, you win, so...

 

The number of battles that are won by completed cap that would not have been won otherwise is pretty low.  (I remember two yesterday, one where the last enemy remaining on Redshire was a french arty that could have cat an moused anything we had for the two remaining minutes, and one on Ensk where we smashed the railyards and I decided to stay in the south cap to make them dig me out, they didn't manage it but were dominating the city team hard, so it would have been a slaughter if we hadn't capped)

 

but that was two out of maybe 40 battles, and that's a pretty high incidence of winning by cap when the outcome of the battle was actually in doubt.  Usually I could play that many battles and it would never happen.  Most cap victories require the capping team to have such comprehensive map control and vehicle advantage that they could easily win by hunting down the few stragglers and gangbanging them.

 

So, barring some extreme outliers, if you're winning by cap it's because you already won.

Share this post


Link to post
Share on other sites

i'm really horrible at statistics so completely lost what is actually happening. So just trying to get a a grasp. The idea is too give tanks individual value's and see how good you do based on this value. (like noobmeter pr does?) and you're now on the point on how to determine these values? or am i completely off? 

Share this post


Link to post
Share on other sites

Pretty much. Noobmeter takes number of games per each tank you have played and then uses that to estimate what damage and winrate you "should have" and compares your real overall winrate and damage to that. The problem with that is that winrate is the most variable stat of all, where it can vary over 20% on the same tank played with the same skill simply by platooning or not.

 

We are trying to do something similar, but with damage and kills, and maybe also def, spot, etc. I have used the vbaddict database to try to find comparable values for each of these stats for each tank. Now we have to try and put everything together into a formula.

Share this post


Link to post
Share on other sites

AH ty :) Great feeling when you feel you almost understand something and you're actually right ;) glad i am able to actually follow this topic again.

Share this post


Link to post
Share on other sites

A few questions please:

 

1.

Is there already a spreadsheet available like the previous WN8 Excel sheet where we stat whores can test our values?

 

2.

Is it possible that NOT playing a certain tank can alter my WN8 score?

I´ll give an example to show what I mean:

Let´s say I have 500 battles in an overpowered T29 with a WN7 of 1800. Now WN8 gets released and due to normalizing its OPness my value drops to 1600. I ragequit and play other tanks. Suddenly T29 receives a massive nerf and its (vbaddict) op value drops. Would this raise my personal WN8 of that particular tank (and overall WN8) without playing it again?

Sorry if this is a dumb question...

Share this post


Link to post
Share on other sites

Values for tanks should remain constant unless a tank recieves a huge buff, or nerf, and they will not be updated often.

 

When new tanks are added, they should simply be fit to the average top for the other tanks in their class/tier combo. They can be fixed later, when there is more data available.

 

Also, we should work out the multiplicative function for expected stats based on tier is going to be for each stat.

Share this post


Link to post
Share on other sites

Okay, I used the vbaddict data to make a player DB, only considered players with more than 1500 battles. I have 16051 players in the database, and it seems to me it is good enough to work with the formula. For the tanks for which we have low amount of data, I used the lowest stat line for same tier/class tank. Also, this data has not been tampered with after calculation of stats, as shnog perhaps suggested increasing low tier tank stats by (20%-tier*2%), with a 0% increase for tier 10. I didn´t do that, this data is absolutely clean. So in thoery with this formula, an outstanding player should get the same scores playing tier 1s as playing tier 10s. 

 

Whether or not low tier play has to be somehow "penalized" in order to combat sealclubbing, is a discussion for another day, and I welcome any feedback and opinions on that.

 

First some correlation analysis between the different stats. rSTATS are stats which have been normalized by the "expected" STAT for the number of games that player has played on each tank. rFRAGS and rDAMAGE correlate almost perfectly, and they kind of cancel each other out, and both correlate very well with rWINRATE. rDAMAGE seems to be a superior predictor of wins than rFRAGS, which is suprising given that frags is much better than damage. This is obviously explained by the fact that average tier normalization was not working as well as it could have for damage, and we all know it was WN7´s weakspot.

By normalizing per-tank, damage seems to take over as the main stat, with solid statistical backing.

 

The other three stats have poor correlation with rWINRATE, although there is some correlation up to a limit score of around .95 rSPOT and rCAP, and 0.8 rDEF.

 

l417.jpg

 

What I did was use these rSTATS, each multiplied by different multipliers which I left up to an evolutionary algorithm to decide, using a least sum of squares method to adjust the resulting formula to rWINRATE. The limit I did set was for the multipliers of rDAMAGE and rFRAGS to add up to at least 0.8. Also, I set a cap of rDEF to 0.8, rSPOT and rCAP to 0.95. This increased WN8 correlation to rWINRATE greatly. It seems that players above these numbers can vary greatly in wins, but having less than the expected values makes players lose more. 

 

So the final formula looks like this: 1600*(0.24*rFRAGS+0.67*rDAMAGE+0.16*rSPOT(capped at 0.95)+0.08*rDEFENSE(capped at 0. 8)+0.15*rCAP(capped at 0.95)). The formula looks very good. It´s Rsquared correlation to rWINRATE is 0.7697, very close to what WN7´s is (0.82), but remember WN7 has winrate in the formula and WN8 does not. Also, the average WN8 for this group of players is 1270, and I expect the scale to be similar to WN7, although the number of unicum and superunicum players will decrease, since all the "false" superunicums who abused tier 1s, Hotchkiss, M4s, PzIVs, Hellcat, T49,VK3601, etc. for THOUSANDS of games will watch their WN8 be 200-500 points lower than their WN7 depending on how many games they played those tanks.

 

 

I will clean up the spreadsheet and post it so everyone else can play with it, and possibly come up with different solutions/formulas.

 

 

 

When new tanks are added, they should simply be fit to the average top for the other tanks in their class/tier combo. They can be fixed later, when there is more data available.

 

Also, we should work out the multiplicative function for expected stats based on tier is going to be for each stat.

 

 

For sure tanks without data should eb considered as either the average or the minimum for the tier/class combo. I used minimum due to average possibly being unfair for some tanks. I prefer to err on the safe side.

 

Lastly, the tier multiplicative function should be discussed in length before we try to implement it. I tried it on the data to see how it worked, and I think something closer to 1.07-.01*tier would be better.

Share this post


Link to post
Share on other sites

Rethinking this, I will make a new thread to get people´s attention, we need some feedback, and it seems people don´t look at this thread very much, shnog you scared them away with all your formulas! :D :D :D

 

Heres the spreadsheet:

 

https://www.dropbox.com/s/5348m9gnrecx9tn/vbaplayerDB.xlsx

Share this post


Link to post
Share on other sites

Wait a minute... we now have to adjust for cross-correlations.

 

Could you post the plots and R2 between the rSTATs now?

 

I could if I work out how to do that with excel.

 

Just to fine-tune it a little bit more before release.

Share this post


Link to post
Share on other sites

Pls disregard cap, its BS and most people agree with it.

I dont get the logic behind suddenly putting that nonsense in. I mean before WN-rating everyone and his dog whas farming it and some still are, especially players with above avg stats. Taking data from vbaddict and including cap + trying to create a correlation between cap and win is like a self fulfilling prophecy.

Share this post


Link to post
Share on other sites

Praetor, how do I do this? I want to try tying the stats together like I explained and also adjust for cross-correlations, but I don't know how to do it in excel.

 

Pls disregard cap, its BS and most people agree with it.

I dont get the logic behind suddenly putting that nonsense in. I mean before WN-rating everyone and his dog whas farming it and some still are, especially players with above avg stats. Taking data from vbaddict and including cap + trying to create a correlation between cap and win is like a self fulfilling prophecy.

 

It is not creating a correlation. It is observing a correlation. Also:

A self-fulfilling prophecy is a prediction that directly or indirectly causes itself to become true...

Share this post


Link to post
Share on other sites

Folter, data is pretty consistent. I agree about cap, but remember it is capped at a low value. Farming it will not increase your WN8. It is merely there to "punish" players with extremely low values. Same for defense and spots, suiscouting will not increase WN8. Also, this rCAP is much better than raw cap value. It gives us comparable cap values between a maus player, arty players and a Batchat player.

 

 

Your opinion is valued, I am still playing around with everything, this is just what came out without applying any kind of decisions or choices to filter data. I will now check what the average contribution of each stat is to overall WN8. I will also check average rSTATS and report that. I was expecting the evolutionary algorithm to throw cap value into the ground, close to 0, as what hapened for raw cap values, but for some reason that did not happen... the data is telling us something. It is worth looking into.

Share this post


Link to post
Share on other sites

Checked out some stuff... looked for some players which are on vbaddict and we know are aweomse, like Heavytwenty, jacq132, and they get roughly  equal WN8 as compared to their WN7 , while many "fake" unicums get lower or much lower WN8 scores as compared to their WN7 scores (the top WN7 player on vbaddict is a guy with 7000 games on Hellcat, goes from 3050 WN7 to 2550 WN8).

 

Decided to lower the cap on rDEF and rCAP to 0.7 since the 0.9 values corresponed to about 1.9 cap and 1.4 defense, and that was quite high. 0.7 corresponds to roughly 1.4 cap and 1.2 defense. Raised cap on rSPOTS to 1.2. Also, here are the percentiles for all the stats and rSTATS. 

 

jm4h.jpg

 

 

 

As you can see, CAP, SPOT and DEF are used to mainly tell apart 900-1440 WN8 players. Above that, WN8 relies on frags and damage only. This i what came out of correlation analysis and the evolutionary algorythm, and it makes perfect sense, as it is quite obvious trying to tell apart players in the 1700-2300 range using cap, def and spot, is pretty ridiculous.

 

In average, frags account for 18% of total score, damage for 56%, spots for 15%, defense 5% and cap 5%.

 

Formula is now:

1690*(0.24*rFRAGS+0.71*rDAMAGE+0.16*rSPOT(capped at 1.2)+0.08*rDEFENSE(capped at 0.7) +0.08*rCAP(capped at 0.7)).

 

Share this post


Link to post
Share on other sites

I will also check average rSTATS and report that.

 

I meant things like rDMG vs rDEF in the same way that you did rDMG vs rFRAGS.

 

I also think rWINRATE should be redefined. It should compare the stats to the average expected win rate in each tank and the average standard deviation of win rate in those tanks.

 

For example, if Tank A has an average win rate of 40% and a standard deviation in win rate of 10% and Tank B is 60% vs 5%, and a player has 200 battles in Tank A and 50 battles in Tank B, then it should be rSTAT vs (WR - ((200 * 40% + 50 * 60%) / 250)) / ((200 * 10% + 50 * 5%) / 250).

 

EDIT: There should be a win rate part in WN8, and it should use this kind of z-score. Specifically, a smooth function of this z-score.

Share this post


Link to post
Share on other sites

Many, many MANY people were against using winrate in the formula. And I agree partially. It is the only stat that can vary 15% on the same tank, played by the same player, just with or without platoon. I am not comfortable including it.

 

Now that we can use rDMG and rFRAGS instead of damage an frags, I don´t really think we need it. It was used in WN7 as a proxy to measure intangibles, but the truth is, the difference in winrate between a unicum solo (60%) and a good player solo (55%) is much smaller than the difference in winrate for a player with or without platoon (55% vs 70%, 60% vs 80%, 65% vs 90%, etc.)

 

Ill get back to you with those comparisons.

Share this post


Link to post
Share on other sites

Many, many MANY people were against using winrate in the formula. And I agree partially. It is the only stat that can vary 15% on the same tank, played by the same player, just with or without platoon. I am not comfortable including it.

 

Now that we can use rDMG and rFRAGS instead of damage an frags, I don´t really think we need it. It was used in WN7 as a proxy to measure intangibles, but the truth is, the difference in winrate between a unicum solo (60%) and a good player solo (55%) is much smaller than the difference in winrate for a player with or without platoon (55% vs 70%, 60% vs 80%, 65% vs 90%, etc.)

 

Ill get back to you with those comparisons.

 

Yeah, I don't like using win rate either. :)

 

What will the new colour scale be? I suppose we can start discussion on this now.

 

I say this:

 

0% ~ 10% black

10% ~ 20% brown

20% ~ 30% red

30% ~ 40% orange

40% ~ 50% yellow

50% ~ 65% light green

65% ~ 85% green

50% ~ 99% blue

99% ~ 99.9% purple

99.9% ~ 100% deep purple

 

(or whatever the hell the current percentile system is; slipped my mind a bit)

 

EDIT: I now realise that the evolutionary algorithm already makes up for the cross-correlations. :P

Share this post


Link to post
Share on other sites

  • Recently Browsing   0 members

    No registered users viewing this page.