Jump to content
RichardNixon

WN9 candidate prototype

Recommended Posts

Q: is the cap 300 per the wiki text: 'Create an array of Per-tank WN9 values and battle counts, capped to 300'

or is it 200 + total battle count / 50.......like it says in the code in the wiki?

Big difference; with 28000 battles I'd be capped at 760 rather than 300 per tank

Link to post
Share on other sites
13 hours ago, Lowlife555 said:

so my 1100 battle count in Löwe, which I played 3-5 years ago, will be counted? *sigh*

Only 26.5 of them by the current method. It's your borderline tank at 65%.

7 hours ago, Lowlife555 said:

I tought WN9 was supposed to be about recent stats, not tanks we played years ago....

How is that calculated?

Is 10 games in a tier 2 tank counted for instance?

Recent WN9 recent is recent. Account WN9 is pseudo-recent: It picks your better-played tanks, which are usually the most recent but not always.

10 games in a tier 2 tank will only count if you played it well. Which would be unusual.

5 hours ago, Gryphon_ said:

Q: is the cap 300 per the wiki text: 'Create an array of Per-tank WN9 values and battle counts, capped to 300'

or is it 200 + total battle count / 50.......like it says in the code in the wiki?

Code is currently correct, although it may change again shortly

Link to post
Share on other sites
6 hours ago, Esprit said:

Would you add KR server to http://jaj22.org.uk/wotstats.html  for Korean players? :D

KR has a different system for obtaining application IDs, and I forgot what it was. IIRC it involved asking someone, which means I'd have to be in an unusually good mood.

Link to post
Share on other sites

Ok, I ran a global test on how the caps would affect active EU players of various skill and experience. I looked specifically at how many battles people already had in the tanks they were playing, and how often they were playing them. Results were a bit surprising:

1. At all experience levels beyond ~7k battles, the current cap level (200 + battles/50) runs only 15-25% above the average played tank. To put it another way, about 40% of battles played would be cap-affected. That's rather higher than what I was aiming for.

2. At all experience levels, there's a substantial chunk of players playing battles in tanks with much higher battle counts. Even with the caps doubled, there would still be some discriminating effect.

3. At all experience levels, the tier 10 tanks played tend to have around twice as many battles on them as tier 9s, but fewer than tier 5-8 tanks. This is true even with premiums excluded, and also with the sample filtered down to good players (who can pay their way in tier 10). Apparently running mid-tier "keeper" tanks is extremely common.

 

Anyway, this suggests two options:

A. Increase the cap globally to 200 + battles/20 or battles/25. This is generous to people playing keeper tanks, or who simply prefer to hang around in the mid tiers, but also low enough to have little effect on most heavy tier 10 players.

B. Set the cap to 200 + battles/30 or battles/40, but double it for tier 10. This fits the "grinder" pattern better but slows down the players with mid-tier keepers. Increases code complexity.

Having a tier-variable cap may feel like a value judgement, but tier 10 is a reasonable special case because you can't move further down the line even if you wanted to.

Link to post
Share on other sites
3 hours ago, RichardNixon said:

B. Set the cap to 200 + battles/30 or battles/40, but double it for tier 10. This fits the "grinder" pattern better but slows down the players with mid-tier keepers. Increases code complexity.

Having a tier-variable cap may feel like a value judgement, but tier 10 is a reasonable special case because you can't move further down the line even if you wanted to.

I'm a huge fan of doubling the T10 cap. Most skilled players agree that high tier play is more representative of skill and also spend most of their games there. 

Link to post
Share on other sites

Something along suggestion B.

 

I can't see how one can "pad" tier 10 - which I assume is much of the reasoning behind wn9.

 

Generally I'd like to have no cap on tier 10 and somehow diverse cap that is based on tiers. (High tier have higher caps)

Link to post
Share on other sites

Low tier cap sounds good and I hope it will be implemented. At low tier, you play with high skill crews and top consumables vs people with stock tanks. Unfair. At tier 10, you play vs people who usually have the same set-up as you.

I would suggest an uncapped tier 10 and a different cap on low tiers... it can even be as low as 25*tier number, or the number of battles a tank is usually played to unlock the next tank.

Link to post
Share on other sites

I like the idea of the per-tank cap being a function of tier, such as tier * 100; that way you only cap out if you play your loltracktor more than 100 battles or your 62a more than 1000 - which seems about right to me.

Even Tier 10 should have a per tank cap of sorts else you pander to those who have one tier ten and play many thousands of battles in it (there is a guy out there with 18000 battles in a Maus, ffs, but the real target here are those guys with 2000+ battles in the old Tier II T18 (CIA clan lol)

 

Link to post
Share on other sites

I have increased cap using developer tools (in function CalcWN9Account2, line 561) and I was surprised because my overall Account wn9 dropped.

This is strange since it should be increased as more "better" games were taken into calculations. Am I missing something in algorithm?

Link to post
Share on other sites
On 12/07/2016 at 8:13 PM, Vasth said:

This is strange since it should be increased as more "better" games were taken into calculations. Am I missing something in algorithm?

Your high-battle-count tanks are not especially well played. As the bottom 35% are discarded, uncapping the battle count of tanks in the middle third will decrease your account WN9.

This is likely to be a common pattern for players who use a handful of tanks for credit-grinding.

 

On 12/07/2016 at 5:10 PM, Gryphon_ said:

I like the idea of the per-tank cap being a function of tier, such as tier * 100; that way you only cap out if you play your loltracktor more than 100 battles or your 62a more than 1000 - which seems about right to me.

As in the previous post, players appear to stack battles on tanks quite similarly on tiers 5-8, so there would need to be some other guiding logic. I've been looking at the actual costs of playing new tanks:

1. Relative to earned XP, the practical "startup penalty" for non-prem tanks is quite similar at all tiers from 1-9. It's a bit lower at tier 6-7 than at 4-5, which suggests that crew training is the major factor for lower tiers.

2. Medium and especially light tanks generally have a much lower startup penalty. They also have better players, so this may be due to superior crew management & premium/gold use.

3. Given a typical startup, it takes around 400k XP to get to the point where you're padding WN9 on a tank overall. This rises to ~700k XP from tier 6 to 9, mostly because the crews don't improve as fast with XP. However, average XP per battle rises at a similar rate, so the battle count is flat.

The startup penalty is skill-adjusted, but not for management variations. Lower-tier results are based on inferior players, and they're probably less likely to skip stock grinds and buy good crews. The skill difference between tiers 5 and 9 is around 20-30% in the sample.

The argument comes down to whether you balance the caps primarily for the players who are actually playing the tiers (=> pretty flat 1-9), or worry about giving better players an additional incentive to shift good crews into tier 5 tanks and pad those up for a few percent.

 

Leaning towards tier*(40+battles/200) at the moment. That's a 1400 battle tier 10 cap for 20k battle players. Quadratic tier is an option but it's pretty mean on the f2p players.

 

Link to post
Share on other sites

Considered reducing the caps a bit, but light touch is ok. tier*(40+battles/200) has been pushed to the site, along with a number of minor fixes. Details here:

http://jaj22.org.uk/news.html

If no-one complains too much in the next few days then I'll probably ship it.

Link to post
Share on other sites

Had a more detailed look at the expected values (Ru tier 10 meds, tier 9 and 10 TDs and heavys), some middle tier mediums and the expected stats seem fine, both in tier / class and between tier / class, when i look at the super unicum scale, the damage is roughly what i consider good, perhaps a bit low for middle tier tanks, and a bit high for some high tier tanks, but all in all it makes sense

Also compared to some players, and Wn9 seems to match fairly well, player who i consider equal get equal wn9, who are better get higher, and who are a bit worse / 1 tank only players get it lower.

it look all in all good to me :)

Link to post
Share on other sites

I think by changing that color scales you skewed a bit the reason for the new metric and the reason of removing the 35% worst tanks. Since you moved the scales the same number of people that were purple are still purple and the same number that were blue are still blue. So the 10 is still considered 10 without the 35% and the 8 is still considered 8. Wasn't the purpose to get a more accurate reading on player's current skill? if all the 10s remain 10s and all the 8s remain 8s then what's the difference with wn8?

Link to post
Share on other sites
4 hours ago, kariverson said:

I think by changing that color scales you skewed a bit the reason for the new metric and the reason of removing the 35% worst tanks. Since you moved the scales the same number of people that were purple are still purple and the same number that were blue are still blue. So the 10 is still considered 10 without the 35% and the 8 is still considered 8. Wasn't the purpose to get a more accurate reading on player's current skill?

The missing link is that far fewer players have purple recent WN9 than purple recent WN8. Your stats as an illustration:

Overall / recent WN8 (v26): 2126 / 2861 (+35%)

Account / recent WN9: 874 / 904 (+3%)

This follows directly from the purple account WN9 = purple overall WN8 proportion rule and account WN9 being more representive of current skill.

Link to post
Share on other sites

One thing I noticed, is that playing arty well gives you pretty insane WN9, at least compared to WN8. Is this intentional? Checked one friend, although he is probably one of the best arty players on EU, but still, that is disgusting arty stat padding :)  Table is a bit broken, but basically it's blue to low purple WN8, but high deep purple WN9. Pls no metric to encourage clicking. :)

Tank name
downarrow12.gif
Tier
downarrow12.gif
Class
downarrow12.gif
Battles
downarrow12.gif
Winrate
downarrow12.gif
TEFF
downarrow12.gif
WN8
downarrow12.gif
WN9 maxhist
downarrow12.gif
WN9 current
downarrow12.gif
rWin2
downarrow12.gif
rDmg2
downarrow12.gif
rFrag2
downarrow12.gif
rSpot2
downarrow12.gif
rDef2
downarrow12.gif
ConquerorGC 10 SPG 504 62.5% 1806 2241 1002.5 1119.2 1.88 1.76 1.42 0.85 1.31
Obj. 261 10 SPG 1118 62.3% 2007 2633 1157.5 1157.5 1.86 1.84 1.53 0.92 1.14
T92 10 SPG 383 65.0% 2056 2694 1134.7 1271.5 2.12 2.03 1.52 0.94 1.04
Link to post
Share on other sites
10 hours ago, HS5 said:

One thing I noticed, is that playing arty well gives you pretty insane WN9, at least compared to WN8. Is this intentional? Checked one friend, although he is probably one of the best arty players on EU, but still, that is disgusting arty stat padding :)

Meh. He has higher WN9 in the LTTB. 1263 WN9 in the FV3805, but only over a low battle count.

By design, the vast majority of players will get slightly lower WN9 in arty than other classes. This adjustment did require some assumptions, but if an "arty specialist" still has a mix of classes at the top then it's working as intended.

WN8 was extremely harsh on high-skill arty players (by accident rather than design), although bad players could pad with it.

Link to post
Share on other sites

@RichardNixon

From WN9 experimental stats page, describing WG's PR:

Quote

WG's personal rating, read directly from the API. 8800 is approximately equal to WN8 unicum.

Is this accurate / up-to-date? I always assumed most unicums were at least borderline 10k.

Link to post
Share on other sites
19 minutes ago, NThirtyTwo said:

Is this accurate / up-to-date? I always assumed most unicums were at least borderline 10k.

I think that test was based on unfiltered NA accounts. The WG-PR distribution depends on the server (average tier, reroll percentage), and whether you include low-battle-count and dormant pre-9.0 accounts.

Using the same filters as the WN8 vs account WN9 test (30 day activity, 2k battles), unicum WN8 (v26) is currently at ~9340 WG-PR on EU. With the same filters, the mark is closer to 9200 on NA.

Link to post
Share on other sites

PR is heavily influenced by team games (on top of battle counts).

Maybe less NA unicums play team games (and more reroll)?

Link to post
Share on other sites
On 17/07/2016 at 8:25 AM, HS5 said:

One thing I noticed, is that playing arty well gives you pretty insane WN9, at least compared to WN8. Is this intentional? Checked one friend, although he is probably one of the best arty players on EU, but still, that is disgusting arty stat padding :)  Table is a bit broken, but basically it's blue to low purple WN8, but high deep purple WN9. Pls no metric to encourage clicking. :)

Tank name
downarrow12.gif
Tier
downarrow12.gif
Class
downarrow12.gif
Battles
downarrow12.gif
Winrate
downarrow12.gif
TEFF
downarrow12.gif
WN8
downarrow12.gif
WN9 maxhist
downarrow12.gif
WN9 current
downarrow12.gif
rWin2
downarrow12.gif
rDmg2
downarrow12.gif
rFrag2
downarrow12.gif
rSpot2
downarrow12.gif
rDef2
downarrow12.gif
ConquerorGC 10 SPG 504 62.5% 1806 2241 1002.5 1119.2 1.88 1.76 1.42 0.85 1.31
Obj. 261 10 SPG 1118 62.3% 2007 2633 1157.5 1157.5 1.86 1.84 1.53 0.92 1.14
T92 10 SPG 383 65.0% 2056 2694 1134.7 1271.5 2.12 2.03 1.52 0.94 1.04

2241wn8 = 1119wn9 ? 

Holly crap. 

Link to post
Share on other sites

Could you explain how this capping works please on a case of t8 an t10 as most used tiers for keeper tanks?

I have 34% of my games in t10s, where 5 of them have > 1k games played, 2 of them ~2k games each.

My most played tank is my 1st t10, m48:

Tank name
downarrow12.gif
Tier
downarrow12.gif
Class
downarrow12.gif
Battles
downarrow12.gif
Winrate
downarrow12.gif
TEFF
downarrow12.gif
WN8
downarrow12.gif
WN9 maxhist
downarrow12.gif
WN9 current
downarrow12.gif
rWin2
downarrow12.gif
rDmg2
downarrow12.gif
rFrag2
downarrow12.gif
rSpot2
downarrow12.gif
rDef2
downarrow12.gif
                           
M48 Patton 10 MT 2133 62.9% 1536 3331 903.3 903.3 1.60 1.34 1.50 1.33 1.45

From my last ~5k games played (~1.5 years), approx. 460 of them are in m48, from last ~1.5k (7 month) ~200, according to noobmeter.

In 1st table on "experimental stats" tab, where 1455 of my recent games are counted, I have 3.5k wn8 / 971 wn9 but in the tables below its ~920 wn9.

PR, ER, WN7 are showing me around 900 wn9, WGR and WN8 over 1k wn9, while WN9 is 920?

Why do you need approx. 4k wn8 in t10 meds to get 1k wn9?

 

So, if I played 2k games in 1 single tank, 200 of them in last year, another tank with 3k games, 50 of them in recent year, another tank with 700 games and 500 in recent year, what exactly will be counted?

Link to post
Share on other sites
7 hours ago, lazydot said:

Why do you need approx. 4k wn8 in t10 meds to get 1k wn9?

I'm not aware of any tier 10 medium that needs 4k WN8 for 1k WN9, although some tier 9s might. In WN9, all classes are similarly difficult to get high values in (except arty, which is slightly tougher), unlike WN8 where mediums and lights were easy, slow tanks were difficult and arty were impossible.

Note that because the scale was set for account WN9 vs overall WN8, per-tank WN9 values will generally come out lower on the colour scale than per-tank WN8 values, much like recent values do. However, this is tier dependent: Tier 10 average WN9 and WN8 come out pretty close on the scale, but lower-tier tanks will be substantially lower.

The overall-based stats in WN9 (per-tank, account) are harsh on buffed tanks because the expected values are based on recent data. If you played a lot of battles in a tank which was then buffed (eg. M48), your per-tank WN9 may be quite bad.

 

8 hours ago, lazydot said:

So, if I played 2k games in 1 single tank, 200 of them in last year, another tank with 3k games, 50 of them in recent year, another tank with 700 games and 500 in recent year, what exactly will be counted?

For recent WN9, whatever was played within the sample period. For account and per-tank WN9, all of them.

8 hours ago, lazydot said:

Could you explain how this capping works please on a case of t8 an t10 as most used tiers for keeper tanks?

In your case, the tier 8 cap is currently 1510 battles, and the tier 10 cap is 1888 battles. This means that your Type 59 battles each have around half the weight of battles in uncapped tanks, while your M48 battles are weighted around 12% lower than battles in uncapped tanks.

Link to post
Share on other sites
10 hours ago, DHP said:

2241wn8 = 1119wn9 ? 

Probably not. The "WN9 current" value only applies to tanks played after they were last nerfed, so most players should be looking at the "WN9 maxhist" value for their CGC.

Account WN9 uses the maxhist values, so there's no reward for playing overpowered tanks before they were nerfed.

Link to post
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.


  • Recently Browsing   0 members

    No registered users viewing this page.

×
×
  • Create New...