CraBeatOff

How Marks of Excellence work

42 posts in this topic

This formula was relayed to Deus by Major Rampage and has been validated by WoTLabs users
 

 

 

 
EMAi = EMAi-1 + K*(DAMAGEi – EMAi-1)
• DAMAGEi = (Dealt by player in battle) + ( maximum damage by tracking or by spotting) - (team-damage caused to allies in battle)
• k = 2 / (N +1)
• N = 100

 

 

CernoAlpha provides an excellent explanation of the formulas. 

 

me = math nerd. 

 

so the Exponential Moving Average (EMA) is used when you dont really want to store the 100 last results. You only want to store one last result. 

 

Lets say your average so far is 1000 damage / game. 

 

Case a) your new game is 2000 damage: So the new average is 1000 + 2/(101)*(2000 - 1000) = 1000 + 19. So the new battle raises your "average" quite a bit. 

 

Case b) your new game is 100 damage: So the new average is 1000+2/101*(100-1000) = 1000-18. So the new battle lowers your average. 

 

This also shows why the formula doesnt work for very few battles (the average for 2 battles at 1000 damage/ game would be 40) 

 
 

Generally, new battles are TWICE as important for the average as they would be if this was a straight average of the last 100 battles (thus the 2 factor in 2/101). As more battles accumulate the older ones start being less and less important. The 2 factor is used to make the average more responsive to changes, so you dont need to wait 100 good games to see an improvement, it is quite a bit more responsive than that. A 100 factor would make the newest battle as important as all your previous battles combined (which would suck if it is 0 game...). So a new battle that is good can balance 2 shitty older battles, and a shitty new battle cancels 2 great ones. 

 

A few great games in a row, without any bad games, would bump the average the most. However, this group of good games would be less and less useful the more you play after that. 

 

cheers

 

CernoAlpha

 

 

 

Original thread is the best stuff is buried around pages 9-11. Go upvote Deus and Cerno there!

Share this post


Link to post
Share on other sites

I'm looking through that thread, and don't see clarification of this...

 

maximum damage by tracking or by spotting

 

Does that indicate that it only considers your overall best spotting game for that part of the damage calculation?  It seems counter-intuitive to do so, but the language is ambiguous and the ramifications are significant.

Share this post


Link to post
Share on other sites

Using an exponential moving average is expected. What I want to know is where they get the percentage value from. It took games at well over 2k DPG to get 3 marks in the Type 59 - I highly doubt 1% of Type 59 players average that high let alone 5%.

Share this post


Link to post
Share on other sites

Using an exponential moving average is expected. What I want to know is where they get the percentage value from. It took games at well over 2k DPG to get 3 marks in the Type 59 - I highly doubt 1% of Type 59 players average that high let alone 5%.

 

Ah yes, from that same thread its been posited (although not substantiated) that they're just using a normal distrib of DAMAGEi (or some other sigma lolololol you know they love their random sigmas) and extrapolating the 65, 85 and 95% points from there, as if it were a true cumulative distribution, which is almost certainly is not. 

 

Post here is a decent theory (which is consistent with their previous stat ability, WG that is) 

Share this post


Link to post
Share on other sites

Using an exponential moving average is expected. What I want to know is where they get the percentage value from. It took games at well over 2k DPG to get 3 marks in the Type 59 - I highly doubt 1% of Type 59 players average that high let alone 5%.

 

 

This

 

When I was still playing I just keeped shaking my head, while wondering how the top 5% of the playerbase where doing > ~2K DpB on the Löwe. Thats fucking BS and contradicts the vbaddict-data ... .

Share this post


Link to post
Share on other sites

I've asked for clarifications on how the benchmarks are set.  I've concluded that nobody at the NA office actually knows.  Ancient RU secret!

 

The theory Crab listed earlier fits the data, however.

Share this post


Link to post
Share on other sites

the marks make no sense [full stop] the 85% and 95% borders are way, way too high to make ANY sense, afterall:

- WN8 is mostly based / scaled by dmg (if not padded)

- Super unicum = +2900 = better as 99.99%

- Mark 3 is more dmg as 95%

- ????

 

So in theory anyone pulling blue (99%) should get mark 3 after lets say 100 games (if no outliers) while people running (dark) puple should (easy) get 3e mark, even with (way) less as 100 games or with a few not so good games...

 

i now played 50 or so games in KV5 and have 2 stars, if this thing scales ``normal`` i should get 3e mark @ 70 or so games, but i somehow think this wont happen....

Share this post


Link to post
Share on other sites

the marks make no sense [full stop] the 85% and 95% borders are way, way too high to make ANY sense, afterall:

- WN8 is mostly based / scaled by dmg (if not padded)

- Super unicum = +2900 = better as 99.99%

- Mark 3 is more dmg as 95%

- ????

 

So in theory anyone pulling blue (99%) should get mark 3 after lets say 100 games (if no outliers) while people running (dark) puple should (easy) get 3e mark, even with (way) less as 100 games or with a few not so good games...

 

i now played 50 or so games in KV5 and have 2 stars, if this thing scales ``normal`` i should get 3e mark @ 70 or so games, but i somehow think this wont happen....

It took me 70 games for the first mark in the bulldog, I feel like somewhere they added a random constant increasing the values per tier, because otherwise I would be getting marks far more often than what I do.

Share this post


Link to post
Share on other sites

The problem is that DPG for most tanks is a left skewed Gamma distribution, not a normal distribution. When you try and fit a gaussian to that, you force the mean too high. If you maintain the same standard deviation, you then get the 95% threshold (slightly under +2 STD) too high.

 

WG, learn probability distributions PLEASE.

Share this post


Link to post
Share on other sites

The problem is that DPG for most tanks is a left skewed Gamma distribution, not a normal distribution. When you try and fit a gaussian to that, you force the mean too high. If you maintain the same standard deviation, you then get the 95% threshold (slightly under +2 STD) too high.

 

WG, learn probability distributions PLEASE.

 

Can you estimate (guess) what actual percentile the WG-85% and WG-95% points correspond to?

Share this post


Link to post
Share on other sites

Can you estimate (guess) what actual percentile the WG-85% and WG-95% points correspond to?

 

On NA some of the very best players don't have 3 MoE in stuff like the CGC (TD42 only has 2) and the STB (camador only has 2), so some of their "95%" are beyond actual possibility/probability.

 

And while I am not a great damage farmer, I am readily in the top 1% of players but have to work very hard to get 2 and 3 MoE on some higher tier tanks. Its basically a function of how good the average active player is, and whether you can exceed that sufficiently. 

Share this post


Link to post
Share on other sites

On NA some of the very best players don't have 3 MoE in stuff like the CGC (TD42 only has 2) and the STB (camador only has 2), so some of their "95%" are beyond actual possibility/probability.

 

And while I am not a great damage farmer, I am readily in the top 1% of players but have to work very hard to get 2 and 3 MoE on some higher tier tanks. Its basically a function of how good the average active player is, and whether you can exceed that sufficiently. 

 

This ^, tanks were joe average is terrible are easy to get moe, tanks were joe average is good in are next to impossible (im 99.99% sure nobody will ever get 3 moe on maus for example, due to armor / hp / mobility the average player is fairly good with it (see also global winratio) while at at the same time those same traits make it impossible for a good player to excel, mark 2 is (i assume) already ridiculous hard on maus....

 

ps: T62 should be ``easy``, expected stats for T62 are also ridiculous low afterall (in regards to wn8)

Share this post


Link to post
Share on other sites

Yeah, I've given up try to get 3 marks on the B-C, I need to improve much more as a player before that will happen. Despite average 3.2k dmg and 800 assist, I am still only around 87% on a good day.

Share this post


Link to post
Share on other sites

For the beginning, I would like to apologise for all stylistic and spelling mistakes (I haven't write in english for at least 4 years). 

 

 

 

 

I want to ask you guys:

 

Is it confirmed that Marks of Excellence have such formula?

 

EMAi = EMAi-1 + K*(DAMAGEi – EMAi-1)

• DAMAGEi = (Dealt by player in battle) + ( maximum damage by tracking or by spotting) - (team-damage caused to allies in battle)
• k = 2 / (N +1)
• N = 100

 

 

I'm not talking about is it or is it not a EWMA (in USA is EMA) cause it is obvious, that EWMA is the best formula for this. 

I'm rather sceptic to the period of '100' in it.

 

We have to realize that the sum of weightes in last 100 battles for EWMA100 is 'just' 0,865. So that 'historical games' (games from 101 to infinity) have a real influence on our badges.

 

Example:

Player A has exactly 1 '000 dmg from every 100 battles. His EWMA100 = 864.7.
Player B has exactly 1 '000 dmg from every 250 battles. His EWMA100 = 993.3.
Player C has exactly 871 dmg from every 250 battles. His EWMA100 = 865.11.
 
As can we see, 'history battles' has great influance on the EWMA value.
And now we can see that what we can read in WOT: "Only the 100 recent battles are counted" is far far faaaaar away for being a truth..
 
Therefore I suspect that Minsk made EWMA with shorted period for example EWMA65 for which the sum of weightes for 100 battles will be 0.954. Of course, there will be differences in values of EWMA65 for 100 battles. However it won't be so great (for the players from example it will be: 953,9 ; 999,5 ; 870,6).
Of course, if we will consider shorter period for EWMA the sum of weightes for 100 battles will be closer and closer to 0,(9) but there is another thing. For example the sum of weightes of EWMA30 for 10 battles is 0,487 (almost 50% of the mark procent is calculate from last 10 battles - this is to much and of course it's not what we see in the game).
 
So, once again: Is it confirmed that Marks of Excellence have EWMA100 formula not any other period?

Share this post


Link to post
Share on other sites

Ah, this explains why most of my double MOEs and my solo triple (crusader) came on light tanks.  

This would be a nice thing to fix...the better the system, the more meaningful it would be.  

Share this post


Link to post
Share on other sites

Example:

Player A has exactly 1 '000 dmg from every 100 battles. His EWMA100 = 864.7.
Player B has exactly 1 '000 dmg from every 250 battles. His EWMA100 = 993.3.
Player C has exactly 871 dmg from every 250 battles. His EWMA100 = 865.11.
 
As can we see, 'history battles' has great influance on the EWMA value.
And now we can see that what we can read in WOT: "Only the 100 recent battles are counted" is far far faaaaar away for being a truth..
 
Therefore I suspect that Minsk made EWMA with shorted period for example EWMA65 for which the sum of weightes for 100 battles will be 0.954. Of course, there will be differences in values of EWMA65 for 100 battles. However it won't be so great (for the players from example it will be: 953,9 ; 999,5 ; 870,6).
Of course, if we will consider shorter period for EWMA the sum of weightes for 100 battles will be closer and closer to 0,(9) but there is another thing. For example the sum of weightes of EWMA30 for 10 battles is 0,487 (almost 50% of the mark procent is calculate from last 10 battles - this is to much and of course it's not what we see in the game).
 
So, once again: Is it confirmed that Marks of Excellence have EWMA100 formula not any other period?

 

Nice work and well written.  Unfortunately there is no way to actually confirm the EWMA, because the only data that we have is the percentage target that WG gives us.  Based on community activity even that number is wrong, because it is probably a comparison of personal EWMA to population actual values over that time period.

 

Given past WG behavior, it is very likely that they chose EWMA(100) without doing any of the analysis that you provided, in which case it probably has the drawbacks that you noted.  Par for the course.

Share this post


Link to post
Share on other sites

So does anyone know where we could get our hands on a large collection of average values for a given vehicle? Say, the average IS-3 values for 10,000 players? Might be possible to compare to the 65/85/95 required values reported by VBAddict and estimate how Wargaming is coming up with the requirements. Certainly wouldn't be completely accurate, but might be an interesting project.

Share this post


Link to post
Share on other sites

So does anyone know where we could get our hands on a large collection of average values for a given vehicle? Say, the average IS-3 values for 10,000 players? Might be possible to compare to the 65/85/95 required values reported by VBAddict and estimate how Wargaming is coming up with the requirements. Certainly wouldn't be completely accurate, but might be an interesting project.

 

I won't help do this, but I'll lay a wager, say 1500g, that they're computing 1 std dev and then extrapolating the 2nd and 3rd from there, despite the fact that damage isn't even close to normally distributed :-D

Share this post


Link to post
Share on other sites

For the love of god. Read the fucking descriptions. Your past 100 game summed avg versus the HIGHEST FUCKING RESULT of people who played the tank in the past 2 weeks. To get the 3rd mark, you have to average higher than the top 5% of all players managed to get in a SINGLE MATCH in a tank. That's why its a pain, because you are competing against single battle results. Even a pubbie can have a good match sometimes.

Share this post


Link to post
Share on other sites

For the love of god. Read the fucking descriptions. Your past 100 game summed avg versus the HIGHEST FUCKING RESULT of people who played the tank in the past 2 weeks. To get the 3rd mark, you have to average higher than the top 5% of all players managed to get in a SINGLE MATCH in a tank. That's why its a pain, because you are competing against single battle results. Even a pubbie can have a good match sometimes.

Must be new text in the ingame UI? I'll add screens to OP when i can. Relax luna we just missed whenever they added that...

Share this post


Link to post
Share on other sites

Must be new text in the ingame UI? I'll add screens to OP when i can. Relax luna we just missed whenever they added that...

Sorry, I've basically heard this question asked a dozen times by now in various places and I'm tired of it. I'll update my MoE thread and just link that for the future.

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

  • Recently Browsing   0 members

    No registered users viewing this page.