[WN6, WN7 рейтинг]

[sirmax 5,499]

I´m sad you feel that way. I beliee WN6 in default conf would be a HUGE step forward. I do not understand why you have such a problem with very VERY bad players getting 50-150 less points than Efficiency.

Just redraw the categories like Fromshadow said and I think everything would be good.

 

Efficiency is an EXTREMELY flawed metric. Just change it to WN6 and let people get used to the new numbers....

Also, I repeat, if the limit of games is placed at 1000 instead of 200, the numbers will be much more similar.

The problems are:

1. support. MOST of users (99%) will not read about changes, and WILL write to support questions like "why MY stat is lower in that version?", and "better I will use old version because MY stat is higher". Nobody interested in changes in other stat, only personal.

2. ranges. WN6 have maximum effective ranges from -350 to 2600 (no one have velue over this range according real data analisis), and this is the problem. I have tried to replace eff to wn6 in win chance calculation, and result of prediction was worse. The reasons is: a) negative values, b) WN6 is farther from Gaussian distribution then eff.

 

We have 5 options now:

1. Leave WN6 as is.

2. Make smth like WN6''', to make ratings comparable.

3. Make both EFF and WN6 to be close to Gaussian distribution.

4. Use another ranges for WN6 to prevent reflexive comparision of different values. (we provide ranges 00-99 or 000-999). If we use 00-99, the precision of original value will be less in 20 times (wn6 linear effective range is about 2000 points). This is enough for XVM, because we are using cached data. For online data we can use 000-999 range, in this case we lose only every 2 point. Another benefit is that it will use less digits in the numbers, and we willl save place on the screen for another values. Also 2 or 3 digit values are easier to understand.

5. Create new formula. We have very good database for statistic analysis. I want to use data mining for analysis, and I can have very good results. Beleave me, I'm professional in this area. :)

 

I beleave you will understand my position.

[seriych 2,178]

Ну что я могу сказать. Средний рейтинг игроков напрямую зависит от среднего количества боев. Данные для 200-1000-2000 боев:

И что с этим делать?

Построим-ка мы зависимость процента игроков от количества проведенных боев:

Синий график показывает процент игроков, проведших боев больше, чем отложено по горизонтали.

Красный распределение процента игроков по количеству боев. Добавим еще график распределения в области 0-2000, чтобы рассмотреть ее поближе:

Видим, что график вначале круто идет вниз, а затем начинает быть более пологим. То, что он идет круто в низ в районе 200 боев говорит о том, что отбор игроков по базе по критерию 200+ боев был сделан неправильно. Минимальная граница боев должна быть больше, чтобы не зарагивать тех игроков, которые чуть поиграли и бросили.

Теперь вопрос, какую границу выбрать и по какой логике. Можно, например, взять только 50% игроков- это будет как раз примерно 2000+ боев.

@Praetor77,

As I promised, I made graphics for the 1000+ and 2000+ fights. You were right, 200 - it is too small. I'm stupid moron.

But how much is not enough and not many? 2000+? This 50% of the players.

[fakels 1]

нет притирки к шкале eff, есть нормирование по 5 точкам превышения над определенным процентом игроков, к которым все привыкли: сколько красных игроков, сколько оранжевых, сколько желтых и т.д. Задача оставить процент игроков каждого цвета таким же. При этом сделать так, чтобы в этих точках рейтинги eff и wn6 совпадали. Чтобы ты мог знать, повысился твой рейтинг или понизился при переходе на wn6.

забудьте вы уже наконец все что связано с eff.

новая формула = новый диапазон для красных, оранжевых... исходя из этого и получится новый процент игроков красных желтых, зеленых... а то что вы сейчас делаете я по простому и называю как притирка со старой eff.

[seriych 2,178]

забудьте вы уже наконец все что связано с eff.

 

Зачем? Зачем терять возможность сравнения старой формулы с новой? Это примерно как приехать в США и привыкать к милям, не зная, о том, сколько это километров.

 

новый диапазон для красных, оранжевых... исходя из этого и получится новый процент игроков красных желтых, зеленых

Вообще-то диапазон должен быть исходя из процента, а не исходя из "красивых" чисел.

 

 

Кстати, графики wn6 сдвинулись, но и eff сдвинется же.

Edited February 7, 2013 by seriych

[x_e_n_y 5]

Original WN6 formula:

 

(1240-1040/(MIN(TIER,6))^0.164)*FRAGS
+DAMAGE*530/(184*e^(0.24*TIER)+130)
+SPOT*125
+MIN(DEF,2.2)*100
+((185/(0.17+e^((WINRATE-35)*-0.134)))-500)*0.45
+(6-MIN(TIER,6))*-60

 

 

 

Предлагаю модификацию:

 

(1240-1040/(MIN(TIER,6))^0.164)*FRAGS
+DAMAGE*530/(184*e^(0.24*TIER)+130)
+SPOT*100
+MIN(DEF,2.2)*100

+MIN(CPTR,2)*50
+((185/(0.17+e^((WINRATE-35)*-0.134)))-500)*0.45

+((185/(0.17+e^((ALIVE-20)*-0.134)))-500)*0.45
+(6-MIN(TIER,6))*-60

 

Внесённые изменения и их обоснование:
а) уменьшено на 20% влияние первоначального засвета - т.к. первоначальный засвет мало влияет на исход боя и завышает рейтинг ололосветлякам;

б) в формулу возвращены средние очки захвата базы, но ограничены значением (не более 2) и их влияние в 2 раза ниже влияния очков защиты - побеждать захватом можно и нужно и рейтинг должен это оценивать, фармить рейтинг захватом при ограничении в 2 очка не получится;

в) аналогично винрейту рассчитывается % выживаемости (% выживаемости коррелирует со скиллом и винрейтом и компенсирует отсутсвие первначального засвета для арты)

Edited February 7, 2013 by x_e_n_y

[Praetor77 20]

Nice analysis seyrich! I knew 200 was too small, because we tested that. WE decided on 2000, but I guess 1000 will work almost as well, and increase the % of players a lot.

 

Anyways, 2000 battles makes sense because those are the players that are here to stay and are dedicated to the game. You will rarely see players with less than 2000 games in tier 5+ battles, and measuring their skill is complicated because it fluctuates too much during that first part of the learning curve.

 

 

 

Sirmax, back to the analysis, WN6 does not in fact have a normal distribution, in fact it has a gamma distribution, and we know this for some time now.

 

http://en.wikipedia.org/wiki/Gamma_distribution

 

Even without analyzing a sample of players, we knew it would have this kind of distribution since damage and kills have gamma distributions and not normal distributions (we analyed 200000 battles from vbaddict/wot). We tried transforming the data to a normal distribution and then measure each stat in a "mean + x standard deviations" fashion. However this was a VERY complicated formula, and the gain in accuracy seemed to be insignificant (2-5%). 

 

 

This gamma distribution has to be accounted for when calculating chances to win using WN6. This may be why it performs worse than with eff, and I think eff has a closer to normal distribution because a large percentage of players (about 50%) have a huge linear factor that is not influenced by skill, simply by taking the decision to farm at end of game (cap). This increases the value of the shape parameter and decreases the scale parameter, making the curve similar to a normal distribution.

 

That being said, you can see above that with players with a minimum of 2000, WN6 is much closer to Eff %s of players.

 

 

 

Regarding the range, we used a similar range to Eff on purpouse, so interpretation of WN6 scores would be easier. Kind of ironic that you are put off exactly by that reason, in being similar but different at player %s for the server.

 

 

Seyrich, could you compare WN6 vs Eff with the 2000 game player dataset?

 

 

PS: Including survival is redundant, because damage and kills depend VERY much on survival. Secondly, for some reason average spots seem to be a very important part of the WN6 formula, and somehow areable to predict how often a player wins surprisingly well. Remember this is for overall account stats. I guess it has to do with aggresive play or something, plus the fact that you cant farm spots and damage/kills at the same time. If you camp and play very defensively, high damage and kills but  low spot. If you suiscout, lots of spots, but no damage/kills.

Edited February 8, 2013 by Praetor77

[nemoW 84]

EFF formula today has changed. But it's still vulnerable to capture points.

[Praetor77 20]

Interesting. Now that NEW EFFICIENCY scores are lower than old eff, no one will care about WN6 drop in scores... :D

[seriych 2,178]

WN6 does not in fact have a normal distribution, in fact it has a gamma distribution, and we know this for some time now.

 

Thank you said so. This can help me.

 

Now we want to leave all the formulas as it is, but in addition to make our own double-digit scale for all formulas, based on a percentage of superiority. About this:

This is a very rough schedule and will vary. But the the idea is that in the early increment percentage is large, and at the end a little. It's like in both rankings, but I think also in the lower range should not be a knee, which is in the rankings.

 

 

Seyrich, could you compare WN6 vs Eff with the 2000 game player dataset?

 You can look at the second chart in this report:

http://www.koreanrandom.com/forum/topic/2625-wn6-vs-eff-in-default-config-wn6-или-eff-в-конфиге-по-умолчанию/?p=34641

Thick lines- wn6(200, 1000, 2000), fat point-points on the respective charts eff. Thin lines- wn6''' for each value of fighting (the formula is the same as for wn6'''_200). So with the increase in number of fights eff is changing almost as well as wn6. Respectively, and the difference is approximately preserved.

 

 

Regarding the range, we used a similar range to Eff on purpouse, so interpretation of WN6 scores would be easier. Kind of ironic that you are put off exactly by that reason, in being similar but different at player %s for the server.

Just the same range and the other percentages led me to adjust the formula for a single percentage of superiority. If your formula gave values ​​in a different range (eg, 0-999), I might not have thought of doing it.

 

 

 

Even without analyzing a sample of players, we knew it would have this kind of distribution

 

But I do not understand why the players are distributed on the number of fights as we see. I was sure that the graph of the distribution in the form of differential is slightly similar to the normal distribution. Therefore thought that replacing 200 for 2000 will not change anything, because most of the players still have much more than 200 fights. And we can see that from the beginning there is a tendency: the more fights, the less players. There is no maximum at some point, like 1000, 2000 ... Maximum at zero and immediately starting a distribution is strictly decreasing. I do not understand why this is so. I assumed that in the base players in the history of the base, but sirmax denied this, players in 45 days.

 

in fact it has a gamma distribution, and we know this for some time now.

 Wow, they are both so close to the gamma distribution:

thx

Edited February 9, 2013 by seriych

[Praetor77 20]

Yeah, and now all the analysis has to be redone for new Efficiency formula. The new efficiency formula dropped scores for everyone, so I guess, now no reason to complain about WN6 score distribution, lol! :D

 

 

The WN6 score distribution is a gamma distribution for sure. This is when you use ALL the players in the server. As you raise the minimum number of battles considered, it becomes more and more like a normal distribution, or to speak more correctly, this increases the value of the shape parameter and decreases the scale parameter.

 

 

I believe the new Efficiency formula is still pretty bad. Still very farmable by capping and defense. Also suiscouting 12 enemies still gives you 2400 efficiency... O_o

The score of all the best players on the server has dropped with the new effficiency and the scores of pedotankers has actually increased with the new formula!

 

Maybe this is the perfect chance to change to WN6 for default config! New efficiency would drop scores anyways!

Edited February 10, 2013 by Praetor77

[FromShadow 24]

В чем то он прав. Взять того же KEEN_ или d3JaVUE:

 

 

Новая eff их порезала, но не настолько, что бы убрать из топа. wn6 же режет достаточно хорошо.

Edited February 10, 2013 by FromShadow

[seriych 2,178]

The new efficiency formula dropped scores for everyone, so I guess, now no reason to complain about WN6 score distribution, lol! :D

 In reality, this means that, in accordance with the original idea, now both formulas would have to be customized to the old eff. Because the cause was in a familiar old values ​​of the scale, but not in that formula is good.

 

But since the idea some reason almost nobody understands (you also still do not understand, because you write about drop points), leave the formula is. And in addition we make to each formula its double-digit scale based on the same percentage superiority.

 

I do not know what the formula is better, I do not know which will be the default, and I do not know what the scale will be used by default. Questions to sirmax.

Edited February 10, 2013 by seriych

[Praetor77 20]

I thought both things were a problem for you, the scale and the points drop.

 

I don´t really think anyone cares about the scale being a little different.  You can just inform them of this fact by changing the color scale. REALLY, no one cares about the scale, specially if it only changes significantly at low scores 0-1000. I think we all understand the scale problem, but you give it a HUGE importance when the rest of us think it´s a very minor problem.

Edited February 10, 2013 by Praetor77

[seriych 2,178]

@Praetor77,

On the contrary. Everyone, EVERYONE says that they are raised or lowered on this or that rating, not understanding that the scale has changed and fewer rated one can meet a large number of different rating and vice versa. Even you're talking about raising or lowering :-)
 

[Praetor77 20]

Exactly, they care about their own number, and will quickly get used to, and accept,  the change. Also, like I said new eff or WN6, they will drop anyways so...

 

What they don´t care about is a slight change in the % player scale. I don´t see the true advantage in slightly changing the metric to adjust to the old eff scales.

[Edrard 7]

Interesting. Now that NEW EFFICIENCY scores are lower than old eff, no one will care about WN6 drop in scores... :D

 

 

Thats not true, for many players its raised Eff score. Lets say, that the main idia was to lower damage on lower tier and to growth a bit on high tier. For hte most part of the players Eff stayed same or changed for +50\-50

[seriych 2,178]

Thats not true, for many players its raised Eff score. Lets say, that the main idia was to lower damage on lower tier and to growth a bit on high tier. For hte most part of the players Eff stayed same or changed for +50\-50

Естественно для отдельных игроков все может изменяться в обе стороны. Но тенденция такова, что средний рейтинг по новой EFF уменьшился. И чем выше, тем больше среднее уменьшение.

[PROROCK59 0]

Прошу помочь с включением WN6, а еще лучше залейте свой конфиг.  

[Praetor77 20]

The new efficiency formula is even worse than the old one. Yes it gives slightly less cap points, and is more difficult to farm through capping, but it is HEAVILY tier dependant. Two players with same skills, but one with avg tier played 6 and another with avg tier played 9 will get EXTREMELY different efficiency scores.

 

Efficiency was already tier dependant and you got higher score with higher average tier played, but Eff2.0 is even worse.

 

Damage/tier calculation is TERRIBLE. You just get more points for playing higher tier tanks, even if you do worse than at low tiers. This is the points OLDeff NEWeff and WN6 award for doing approximate average damages for each tier (data taken from vbaddict.net). WN6 is pretty balanced from tiers 5-10, while oldefficiency favors high tiers, and NEWeff is just terrible.

 

 

As this following table shows, new efficiency gives you the same points for doing 1120 damage in tier 4, than for doing 1260 in tier 6, 1365 in tier 8 or 1450 in a tier 10 tank.... unbelievable:

 

[seriych 2,178]

@Praetor77,

Look test7, the default is {{xwn}}:

http://code.google.com/p/wot-xvm/downloads/detail?name=xvm-3.3.2-test7.zip&can=2&q=