RichardNixon

WN9 candidate prototype

646 posts in this topic

My scores are:

72.0 actual win rate

71.9% adjusted win rate

0.66 platoon padding

wn8 = 3288

wn9c1 = 2288

wn9c2 = 2276

__________________________________

Per tier is reasonably consistent, with tier 1-3 being much higher (2600-3100).  Its worth noting I almost never play tiers 1-3, less than 1000 total games.

My best padded wn9c1 tank is the S-51.  @[email protected]  My scores in high tier tds are wildly divergent from wn8.  Wn8 I tend to be like 3k at best, but wn9 is very high (2800+).  Stats pattern seems a lot more similar to wn7 (ie, pad in tds and low tiers).

 

Share this post


Link to post
Share on other sites

First off, seeing news on the metric front is amazing. Was a bit worried that no one would attempt to pick back up the torch.

Results and findings are similar to Garbad's here; 2836 WN8 to 2078 WN9c2. Noticed the same push towards TD, lower tier, and arty scores. S-51 was a big winner for both of us. The TD scores have me a bit worried because they just seem to heavily reward damage farmers; things like the roomba, waffen 4, etc. gave me hefty rewards. Whether that's actually a problem with the rating is another thing entirely, because I admit I have a hard time conceptualizing what's actually going on mathematically.

The platoon padding thing is a neat metric in itself; Garbad is a hardcore soloer with that .66. I'm currently at 1.02, which I assume is about average. Also, super props to introducing us to a rating and then immediately giving us a way to look at whats going on with a script.

Edited by Rexxie

Share this post


Link to post
Share on other sites

.17 platoon padding, anti-social tendencies ftw... WN9c2 seems to rate a lot more of my lower tier tanks/light tanks/arty higher (SU-26 being my massive winner), while the wn9c1 rates more of my normal higher tier tanks towards the top.

Share this post


Link to post
Share on other sites

 The TD scores have me a bit worried because they just seem to heavily reward damage farmers; things like the roomba, waffen 4, etc. gave me hefty rewards. Whether that's actually a problem with the rating is another thing entirely, because I admit I have a hard time conceptualizing what's actually going on mathematically.

It should be primarily down to when you played the TDs. The per-tank adjustment is based on recent data, so if you played TDs back when they were good (after 8.6), they'll come out high. Nerfed tanks like the Foch 155, 268, 183 and all the waffles will come out even higher, assuming that you played them pre-nerf.

If anyone played fresh TDs recently and got very high results then that's another matter. I can't tell from your profiles when you played the tanks.

Apparently both you and Garbad did over 2k damage/game in the S-51. Surely that's a pre-8.6 result?

Share this post


Link to post
Share on other sites

Yes. Basically everything on the top end of my WN9 was at some point nerfed from when I last used them, except a select few. Is this just going to be an unavoidable issue? Dealing with tank balance changes has always been a problem for metrics, but I've never seen it so evident than when looking at WN9c. WN8 definitely has humongous problems with LTs and SPGs, but doesn't seem to have nearly the same issue correcting itself for balance changes.

 

edit: should probably mention that neither Garbad or I did 2k+ in the S-51; both of us hover around 1800.

Edited by Rexxie

Share this post


Link to post
Share on other sites

1.17 platoon ratio. RIP. I do feel like I get medals more often than is to be expected though, but maybe just confirmation bias. (I was really solo for my first 10k battles or so)

Overall from 2690 wn8 to 1969 wn9.

E50 wn8 5728, to 3007 wn9 (3x better than avg pubbie? Neat).

I'm guessing that this has removed the whole exponential scaling problem that wn8 had?

I'm also with uwaz, in wn9 it is now my 6th best tank. (I really have no idea why it is rated higher than stuff like 50b, Leo 1, or t-34-3).

Share this post


Link to post
Share on other sites

I used to suck so I am not totally surprised by how bad the values are but negative values seem a bit harsh or is this not intentional 

 

negative wn9.png

Share this post


Link to post
Share on other sites

Yes. Basically everything on the top end of my WN9 was at some point nerfed from when I last used them, except a select few.

I don't see that pattern in mine.  The last 3 tanks I've played (rudy, 131, LTTB) are each in my top 10.  There are also some I haven't played in years (super pershing, marder 2) and others that are middle (T-34-2, Foch, T69).  Maybe I'm the exception, as my stats tend to vary wildly based on how much I tryhard/solo. 

Overall it seems like my scores vary less.  Tanks that I am playing recently (that I know don't have TC in them, etc) range from ~2800 to 2050.  In Wn8 those same tanks range from 6000 to 2300.  Its still just a damage farming metric, but it seems to reward tank selection/tier less than wn8.  We should probably do a field test to find out which tanks are the most paddable.  I could see everyone in bulba flocking back to T67s or E4s, maybe.  OFC if the values move based on what is the FOTM, padders will be constantly rewarded for padding and then ditching a tank once the masses catch up.  Or perhaps padding tanks that are popular with the masses but not padders (kv1).

What are the high end DPG figures you are using for tier 8/9/10 tanks?

 

Share this post


Link to post
Share on other sites

Yes. Basically everything on the top end of my WN9 was at some point nerfed from when I last used them, except a select few. Is this just going to be an unavoidable issue? Dealing with tank balance changes has always been a problem for metrics, but I've never seen it so evident than when looking at WN9c. WN8 definitely has humongous problems with LTs and SPGs, but doesn't seem to have nearly the same issue correcting itself for balance changes.

Using recent data is the only way to get reasonably consistent skill-scaling, which is required to give prevent unicums getting extreme values in lights and mediums. With WN8's method, the skill-scaling results are dominated by bias, so for example you get oddities like the STB coming out as a bad tank for good players.

In theory you could tweak nerfed tanks upwards, and even use a different array for overall and recent values, but because the API doesn't store when a tank was played, you'll never know which value is correct for a particular account. In that sense it's an unavoidable problem: Overall values will always be a compromise.

What WN8 does is slide the expected values gradually towards the current values with each version, although many of the pre-8.6 arty are stuck forever at the old values. This behaviour can be replicated for overall values, if desired.

 

A correction: Apparently there has been no general TD nerf. Not sure how WG failed to do that with all the map, camo and view range changes, but most TDs are doing just as well as they ever did. The specific nerfs to the waffles, Hellcat and tier 10 TDs did have an impact of 10-15% though.

Share this post


Link to post
Share on other sites

Overall it seems like my scores vary less.  Tanks that I am playing recently (that I know don't have TC in them, etc) range from ~2800 to 2050.  In Wn8 those same tanks range from 6000 to 2300.  Its still just a damage farming metric, but it seems to reward tank selection/tier less than wn8.  We should probably do a field test to find out which tanks are the most paddable.  I could see everyone in bulba flocking back to T67s or E4s, maybe.  OFC if the values move based on what is the FOTM, padders will be constantly rewarded for padding and then ditching a tank once the masses catch up.  Or perhaps padding tanks that are popular with the masses but not padders (kv1).

It's a regression method, so it shouldn't be vulnerable to population changes unless the centroid was very low or the data was thin. You'd need to find tanks that either had no good players (mostly below tier 5), or where the good players are doing it wrong. The T67 has by far the highest skill-scaling parameter above tier 4, so that's unlikely to work. The E4 probably gets murdered by artillery these days. Picking tanks popular with newbies worked a lot better in WN8 due to recency bias, although no-one bothered as the high tier meds were even better padders.

Attempting to exploit the formula or the current metagame is probably a better bet. Full tryhard arty ought to work ok, given that a lot of current arty players are just playing for missions. As that's obviously unacceptable, exploiting the increased value of first spots in a damage-padder medium should give you an edge. Feel free to try. I'd be very interested to see how far you can push the spots/game without harming damage padding.

 

I used to suck so I am not totally surprised by how bad the values are but negative values seem a bit harsh or is this not intentional

It should be difficult to get negative values for the WN9 versions unless you have a very low game count in a tank. Essentially if you play really badly and get very bad luck, you can drop below the zero mark, but simply playing badly should put you above zero in the long run.

The negatives are supposed to be capped at zero anyway because they break the sqrts, but I forgot.

Share this post


Link to post
Share on other sites

It's a regression method, so it shouldn't be vulnerable to population changes unless the centroid was very low or the data was thin. You'd need to find tanks that either had no good players (mostly below tier 5), or where the good players are doing it wrong. The T67 has by far the highest skill-scaling parameter above tier 4, so that's unlikely to work. The E4 probably gets murdered by artillery these days. Picking tanks popular with newbies worked a lot better in WN8 due to recency bias, although no-one bothered as the high tier meds were even better padders.

Attempting to exploit the formula or the current metagame is probably a better bet. Full tryhard arty ought to work ok, given that a lot of current arty players are just playing for missions. As that's obviously unacceptable, exploiting the increased value of first spots in a damage-padder medium should give you an edge. Feel free to try. I'd be very interested to see how far you can push the spots/game without harming damage padding.

 

It should be difficult to get negative values for the WN9 versions unless you have a very low game count in a tank. Essentially if you play really badly and get very bad luck, you can drop below the zero mark, but simply playing badly should put you above zero in the long run.

The negatives are supposed to be capped at zero anyway because they break the sqrts, but I forgot.

I'm a terrible damage padder, especially in mediums.  Kewei tops my scores by up to 50% in mediums.  I can narrow it a lot in tds/heavies, but not mediums.

Check my batchat perhaps if you can screen out the clan wars fights.

Use this to see only my puberinos:

https://clantools.us/servers/na/players?id=1000125653&stats=pubs

Share this post


Link to post
Share on other sites

When/if WN9 is released. There really needs to be a standard chart that comes out with it. One that can't be disputed/reworked by people that feel bad because they're on the low side. Preferably one that uses true population percentages, though I understand the problem with trying to get overall percentage data for an entire server before release.

I'm willing to donate processing power/internet connection to run a server-wide check on all players on NA, maybe others later on, once you get a formula you plan to push. Though even at full speed, it would likely take around 2 days per million players.

Share this post


Link to post
Share on other sites

Yes. Basically everything on the top end of my WN9 was at some point nerfed from when I last used them, except a select few. Is this just going to be an unavoidable issue? Dealing with tank balance changes has always been a problem for metrics, but I've never seen it so evident than when looking at WN9c. WN8 definitely has humongous problems with LTs and SPGs, but doesn't seem to have nearly the same issue correcting itself for balance changes.

 

edit: should probably mention that neither Garbad or I did 2k+ in the S-51; both of us hover around 1800.

I agree with Rexxie - while WN8 has problems with LT, SPGs and high tier MTs, WN9 goes totally insane with some older tanks (my M44 is 3K+). WN8 is clearly better in that regard.

 

What happend with your suggestion for a 2-point WN8? Did you follow up that thought? I m  of the opinion that your idea regarding WN8 with 2 points holds much potential, when its comes to WN8s problem with LTs, SPGs and top tier MTs.

 

 

Share this post


Link to post
Share on other sites

Question/idea: for your platoon ratio, how did you handle Radley Walters, Pools and Bölters medals?

 

While I can remember a few occasions where I got a Top Gun in a platoon, I cant really remember getting a Radley Walter or Pools.

Share this post


Link to post
Share on other sites

Wow. Seems like I have to pad my stats with arty now. My SU 26 stats are just insane even though I just drove it for 9 battles and over 4k/5k WN9 o.O

Share this post


Link to post
Share on other sites

My scores are:

  • 64.7% actual win rate
  • 64.4% adjusted win rate
  • 0.89 platoon padding
  • wn8 = 3005
  • wn9c1 = 2198,3
  • wn9c2 = 2116,4

If i sort for wn9c1 i find the results / sorting acceptable, Vk28 is my best tank, followed by T69, 183, bulldog At15, hummel, lowe, Mle 48, some benefit from old (unbeatable) stats, while with others like AT15 and Lowe im simply playing good (or lucky).

Also all tiers and classes are mixed, its not a blob of tier 8 heavys or tier 10 meds as ``leading`` in terms of wn8, its a huge ``chaos``, it seems you can ``pad`` wn9 is any tank or class, not just tier 9 or 10 mediums ^^

If i sort for Wn9c2, tds are getting more dominant (perhaps a bit too much), the tds which are high up my stats list are however all ``deserved``, 100% winratio at valentine at gives 2977 wn9, well, its 100% winratio, same 183, Foch, ISU, at8 E3, my stats in those are simply rly good, so having rly high wn9 is ok.

The only problem is arty, SU-26 is my best tank, yet i hated it and played terrible with it, also hummel is too high. The root is pre 8.6 arty games + aditional nerfs to both hummel and Su-26, making the old stats unbeatable.

I also looked up some other guys and their stats are ``fairly op``, and the pattetn is clear, arty is almost always the best class, the gap is not that massive, but especially players who played it more a few games (so they got actually good at it) get really high numbers.

TL:DR: All in all it looks good, and an improvement over wn8 (due to much harder of padding) however arty gets rewarded too much (which may lead to all bads starting to play arty, to get good numbers, yolo, even the smallest chance of increasing arty due to a new skill metric would be terrible :P

Share this post


Link to post
Share on other sites

however arty gets rewarded too much (which may lead to all bads starting to play arty, to get good numbers, yolo, even the smallest chance of increasing arty due to a new skill metric would be terrible :P

my biggest concern also:eww:

btw:random retarted clicking, no idea about the game in m7 priest: 3917.8 wn9
4500 avg dmg , almost 3 avg kills in 907: 3100.2 wn9

eggsealk.pngmiddlek.png

Edited by veitileiN

Share this post


Link to post
Share on other sites

Oh, come on, this rates as my best tanks with more than 50 battles KV-220 (in which I just do a massive yoloderps and hope the armor holds long enough), S35 (same), and Otsu the sealclubber... (well, and Chaffee above them, but that is about right - Chaffee is love, chaffee is life)

Share this post


Link to post
Share on other sites

 

The only problem is arty, SU-26 is my best tank, yet i hated it and played terrible with it, also hummel is too high. The root is pre 8.6 arty games + aditional nerfs to both hummel and Su-26, making the old stats unbeatable.

I also looked up some other guys and their stats are ``fairly op``, and the pattetn is clear, arty is almost always the best class, the gap is not that massive, but especially players who played it more a few games (so they got actually good at it) get really high numbers.

Yeah, the main issue is that the 8.6 nerfs were massive on some tanks. The SU-26 and M7 Priest had damage output nearly 100% higher before 8.6. Many mid-tiers were ~50% higher.

Recent values are fine. For overall values I can fudge it with a blend, although it's questionable whether it's necessary. I'm not sure what to do with per-tank results. Some options:

  1. Always use the peak values.
  2. Use a blend of pre-nerf and post-nerf values.
  3. Use a heuristic fudge to guess whether the tank was played pre- or post-nerf.
  4. Provide two values.

I'm also fine with putting an arbitrary penalty on artillery, say 10-15%. I don't think anyone wants a lot of artillery in their games, even artillery players. There's a similar argument for TDs, although rather weaker these days.

 

Question/idea: for your platoon ratio, how did you handle Radley Walters, Pools and Bölters medals?

While I can remember a few occasions where I got a Top Gun in a platoon, I cant really remember getting a Radley Walter or Pools.

I got my Pools medal in a platoon. IIRC, only two of my team died. Freak games are freak games. One issue with modelling the platoon ratio is that the medals are too rare, which prevents it working reliably for anti-padders: A lot of them have no Top Guns at all, and if they triple-platoon, they never get close to a CC. Even rarer medals wouldn't help.

It's just a model though, with plenty of known and unknown flaws. There's no source data on platooning to compare against.

 

What happend with your suggestion for a 2-point WN8? Did you follow up that thought? I m  of the opinion that your idea regarding WN8 with 2 points holds much potential, when its comes to WN8s problem with LTs, SPGs and top tier MTs.

It's all intercept + slope based, and so effectively 2-point. You regress player WN9 vs tank WN9 and use the whole line, rather than picking a point on it. You can do this with pseudo-recent data (as used for WN8), but the slopes are very unreliable and often don't correspond with intuitive skill-scaling. With recent data it's much better: Fast tanks come out with high skill-scaling, while SPGs come out very low.

As a side point, there's an assumption in the regression that player "skill" is single-valued. There's some evidence that this isn't strictly true for arty, and so genuinely good arty players (rather than good players playing arty) may find it very easy to get high numbers.

 

Share this post


Link to post
Share on other sites

Oh, come on, this rates as my best tanks with more than 50 battles KV-220 (in which I just do a massive yoloderps and hope the armor holds long enough), S35 (same), and Otsu the sealclubber... (well, and Chaffee above them, but that is about right - Chaffee is love, chaffee is life)

Sounds like optimal KV-220 tactics? The method rates the KV-220 as the most overpowered tier 5 tank for most players, and apparently you're doing better than that. I guess other players aren't yoloderping enough.

Tiers 1-3 are broken by the newbie protection so I wouldn't take too much notice of them.

Share this post


Link to post
Share on other sites

My Scores:

  • W/R: 64.0
  • Adj w/r 65.0
  • Plat ratio: 0.81
  • wn8: 3135.2
  • wn9c1: 2294.2
  • wn9c2: 2246.1

I lose about 50 points from c1 to c2 for enjoying LT's (can you sense the butthurt), and I feel my plat ratio is way too low. I platoon all the damn time, however I only rarely seek out BiA's and I have a shitload of top guns et al. because I actually chase those (play4fun). 

Also, why does my Adj w/r go up compared to flat w/r?

Share this post


Link to post
Share on other sites

 My Scores:

  • W/R: 56.2%
  • Adj w/r: 55.8%
  • Plat ratio: 0.79
  • wn8: 1896.3
  • wn9c1: 1690.1
  • wn9c2: 1617.9

Anyone smarter than me at the math able to tell me what this means?

Edited by Errants
Aside from, "I'm Bad." I know that part.

Share this post


Link to post
Share on other sites

  • Recently Browsing   0 members

    No registered users viewing this page.