Wednesday, March 13, 2013
Special Thank You To wot-replays.org
In a nice little deal with WoT-Replays.org, we have expanded our collection of replays nearly five-fold to nearly 60k. That means that WoT-Replays.org will have their replays backed up for future contingencies, while we get humongous extra set of data!
Friday, February 15, 2013
Up we go!
Since we now have a model, however mediocre it may be, of tier one battles, let's extending it to add in tier two tanks. This shouldn't be too hard, tier one and tier two tanks fight each other a lot, so I went ahead and pulled a data-set of all battles featuring exclusively tier 1 and 2 tanks. That, by the way, is over 8000 battles
The first thing to look at is the Match Maker. First let's look at the distribution of numbers of tier 1s and 2s:
 |
| It's symmetric, which means I didn't mess up this time |
So, that's interesting, but what about the distribution of battles that Tier 1 tanks get? Let's count the number of Tier 2 tanks and see what we find:
 |
| I have more charts where that came from! |
Another interesting item I found, which at first confused me, is that Tier 2 Artillery can get into battles where all the tanks are Tier 2. In all 1308 battles where there are only tier 2 tanks present, this happened only 4 times. The obvious answer is that these are actually battle tier 3 battles which happen to be missing tier 3 tanks, but the amusing part is that it can happen.
To continue on down the road, I'm going to use the tier one model we put together last post to get a first look and see how well it matches our data. If you don't remember, this is what that model looked like:
| Tank | Odds Ratio |
|---|---|
| Leichttraktor |
1.0239
|
| T1 Cunningham | 1 |
| MS-1 | 0.9495 |
| Renault NC-31 | 0.9181 |
| Renault FT-17 | 0.9096 |
| Vickers Med. Mk. 1 | 0.8563 |
Unsurprisingly, this model doesn't perform all that great on the new data, even when limited to battles that include Tier 1s, only correctly predicting 52.2% of the the outcomes, compared to 51.6% for just choosing side 1. So, obviously, it's back to the drawing board to come with a new model, one that takes into account the new data. The model I'm going to try is going just contain the non-premium tanks, so we won't need a baseline.
Quick check to see if this makes any sense and... Hmmmm. The worst tank, by a long shot, is the Med Mk. II, a tier 2! I suppose that's not unsurprising, we've previously said that British low tiers suck. The fact that the Mk. II is actually 3% worse from the Mk. I is also highly amusing. Worse, the bottom three tanks are all British.
At the other end of the spectrum, the obvious is made evident, the 'derp' T18 dominating the low tier battles, followed by the venerable Pz. II. The difference though is quite immense,
The three 'classic' Tier 1 tanks also place higher than quite a number of tier 2 tanks. The LoLTraktor once again tops the list with the T1 and MS-1 close behind, likely due to the number of experienced players running them for fun and credits. In fact, some of the T18's power may derive from similar sources, though the 75mm seems to be the main cause.
But naturally, if we can't predict much, this isn't worth much. So, of the 8453 battles, it's important to note that this model only predicts 4658 of them, just 55.1%. The null model, just guessing a Side 1 Win, can predict 51.8%. That's an improvement sure, but not much.
 |
| At least the Trend is in the right direction |
Obviously, we're showing that the choice of tank is not the biggest factor in whether a team wins. Personally, I expect we could get a lot of extra predictive power from knowing how many battles players have. It's likely that the experience of playing for a long period of time allows you to do much better against the 'noobs'. While I can't exactly vouch for a skill-based match-maker, I think that this is a perfectly good time to suggest that players with less than, say, five-hundred or a thousand battles be segregated into their own matchmaking pool for their own protection and enjoyment of the game. At least for the first couple of tiers.
For the next post I'm going to add in the low tier premiums to see if we can improve our predictive powers. Hopefully this time it won't be a week and a half between posts. After that I am going to try to tackle high tier tanks, since those are what people actually care about. By then 8.4 will likely be out and my data will be woefully out of date, so we'll go forward with obtaining more data and taking a look at how that huge 183 changed the meta game.
For the next post I'm going to add in the low tier premiums to see if we can improve our predictive powers. Hopefully this time it won't be a week and a half between posts. After that I am going to try to tackle high tier tanks, since those are what people actually care about. By then 8.4 will likely be out and my data will be woefully out of date, so we'll go forward with obtaining more data and taking a look at how that huge 183 changed the meta game.
As always, if you enjoyed this post, please help us in our data gathering efforts. It's simple, it's easy and you can find out how just by clicking here.
Friday, February 8, 2013
A Slight Deviation for Damage Stats
Update Notice: Overlord has said that 2.5 Standard Deviations is the cut off. Edits in Bold have been added to reflect this.
As you may know, it ALWAYS feels like you're getting min-damage rolls. And the funny thing is, you're not entirely wrong. A few people have been making posts on the forums and the subreddit talking about the cause of these issues. Since the intent of this blog is to delve into the basic stats the biases of the game, I figured that this would make a good post.
If you understand the way that the Damage and Penetration (and even the aim-deviation) RNG works, it makes a lot of sense that you would see a lot more minimum damage rolls than you would expect. If you're looking for conspiracy theories, you might even think that this is done on purpose.
If you've been paying attention, you know that Damage and Penetration values vary +/-25%. If this was a uniform distribution of values, the chance of getting any damage value would entirely dependent on your mean damage. The probability of each integer damage value X would be, given a mean damage D, simply:
values +/- 25% from the mean. Now, in case you don't know what a normal distribution looks like, Wikipedia provides a nice example of 3:
Obviously, one can pull those values off the standard Z-tables. For each of these cases, we can determine the standard deviations of the distributions of based on each gun's damage. It's a pretty simple equation:
EDIT: It has come to my attention that Overlord has clarified that the +/- 25% is a 2.5 Standard Deviation limit.
Added Damage Distribution Charts: All Y-Axis Scales are the same for easy comparison.
As you may know, it ALWAYS feels like you're getting min-damage rolls. And the funny thing is, you're not entirely wrong. A few people have been making posts on the forums and the subreddit talking about the cause of these issues. Since the intent of this blog is to delve into the basic stats the biases of the game, I figured that this would make a good post.
If you understand the way that the Damage and Penetration (and even the aim-deviation) RNG works, it makes a lot of sense that you would see a lot more minimum damage rolls than you would expect. If you're looking for conspiracy theories, you might even think that this is done on purpose.
If you've been paying attention, you know that Damage and Penetration values vary +/-25%. If this was a uniform distribution of values, the chance of getting any damage value would entirely dependent on your mean damage. The probability of each integer damage value X would be, given a mean damage D, simply:
For a 100 damage gun? P(X) = 1.9% for any damage value between 75 and 125. A 200 penetration gun? P(X) = 0.9% for any damage value between 150 and 250.
Now, that would seem like an easy, logical choice of a random value scheme. In fact, it's actually rather computationally "cheap", which leaves me confused by this next part:
That's not what Wargaming did. Wargaming decided that they wanted to concentrate the damage values around the average, but still provide a degree of randomization. To do this, they decided that the damage (and penetration) values should be pulled from a normal distribution, but limited to a range of
 |
| I certainly didn't make this, but I could have. I think. |
The so-called 'Standard Normal' (doesn't that sound redundant?) distribution is shown in red. It has a Mean of 0 and a Variance of 1. The important thing about this guy is that you can take any other normal distribution and convert it into a Standard Normal distribution with some simple calculations. Further, because of difficulty of integrating an arbitrary normal distribution, there exist standard tables of values, called Z-tables, for calculating what percentage lies under parts of the curve.
Okay, that's all well and good, but what does that tell us about are damage values? I swear I'm getting to it. One additional property about the normal distribution is important: It extends from -∞ to +∞. That means that in order to limit the variation to +/- 25% of the mean, you have to set some hard cut offs. The way Wargaming did this was to say "If we get a value below X, we'll just call it X. If we get a value about Y, we'll call it Y." What this did was to take all of the probability that was below X and give it to X and similarly take all of the probability above Y and give it to Y.
Now I come to a bit of a conundrum, there is a hidden parameter here that we don't know. Where we know the mean damage values for each gun, we don't know what values Wargaming has chosen for Variance (Edit: Overlord has mentioned that it is a 2.5 Standard Deviation Variance). Since Variance determines how flat or pointed the distribution is, we can't say exactly what proportion of hits should be minimum damage. However, by choosing some logical limits, we can give a good idea of what it should look like.
The obvious choices for the +/-25% limits are at the 1, 2 and 3 standard deviation lines.
Okay, that's all well and good, but what does that tell us about are damage values? I swear I'm getting to it. One additional property about the normal distribution is important: It extends from -∞ to +∞. That means that in order to limit the variation to +/- 25% of the mean, you have to set some hard cut offs. The way Wargaming did this was to say "If we get a value below X, we'll just call it X. If we get a value about Y, we'll call it Y." What this did was to take all of the probability that was below X and give it to X and similarly take all of the probability above Y and give it to Y.
 |
| The white areas would be the probabilities rolled into Min and Max values |
The obvious choices for the +/-25% limits are at the 1, 2 and 3 standard deviation lines.
| Cut Off (#SDs) | % Min Value Hits |
|---|---|
1
|
15.8%
|
1.5
|
6.68%
|
2
|
2.2%
|
2.5
|
0.6%
|
3
|
0.1%
|
Obviously, one can pull those values off the standard Z-tables. For each of these cases, we can determine the standard deviations of the distributions of based on each gun's damage. It's a pretty simple equation:
(Mean Damage)/(4*#SDs)= Damage Standard Deviation
There is one final question to be answered, how much more often do these extreme values occur than any other value? To answer this, I'm going to find the ratio of Minimum damage hits to Mean value hits. I'm going to assume that WG rounds fractions to the nearest integer, so the 'Mean' damage hits actually occur in when a value +/-0.5 from the mean is selected.
| Gun Type | Mean Damage | 1 SD | 2 SD | 2.5 SD | 3 SD |
|---|---|---|---|---|---|
76mm M1A1
|
115
|
11.38
|
0.793
|
0.173
|
0.024
|
88mm L/56
|
220
|
21.78
|
1.516
|
0.3315
|
0.0459
|
105mm T5E1
|
320
|
31.68
|
2.206
|
0.48
|
0.0668
|
122mm BL-9
|
390
|
38.61
|
2.688
|
0.588
|
0.0815
|
130mm S-70A
|
550
|
54.46
|
3.791
|
0.822
|
0.1149
|
155mm SA 58 AC
|
850
|
84.177
|
5.8589
|
1.27
|
0.1775
|
170mm PaK46
|
1050
|
103.9
|
7.237
|
1.579
|
0.219
|
EDIT: It has come to my attention that Overlord has clarified that the +/- 25% is a 2.5 Standard Deviation limit.
Added Damage Distribution Charts: All Y-Axis Scales are the same for easy comparison.
 |
| 88mm L/56 Damage Distribution |
 |
| 76mm M1A1 Damage Distribution |
Since the mean value is supposed to be the most likely roll, it will always be more likely than any other roll (by varying amounts). The interesting part though is the trend. As your average damage goes up, these extreme values (Min and Max rolls) increase in proportion.
If we go with the 2 Standard Deviation case , the one I find to be most likely (Particually since I remember WG saying that the aim circle was the 98% line), you'll see that a Jagdpanzer E-100 will see a min or a max damage roll 7.2 (1.58 2.5 SDs) times more often than they will see any other number! More typically armed high tier tanks will see min or max rolls 2-4 times more often (Half to 80% as often) as any other number, and all tanks with average damage greater than ~140 will see minimum and max rolls more often than average rolls.
 |
| 170mm PaK46 Damage Distribution |
 |
| 120mm M58 Damage Distribution |
Keep in mind though, for the 2 SD case, that Minimum and Maximum rolls will still only make up 4.4% of the total rolls we see. For the 2.5 SD case, they would be 1.2% of all rolls. But they are more likely than any OTHER most other specific values, especially those a bit removed from the mean, so they will seem to show up a lot (plus, how many people remember how often they hit for 159 damage? There is a bit of a confirmation bias effect going on here too).
As for Max damage rolls, there is one additional reason we see less of them than Min rolls: Low Health Tanks. When you hit a tank that has hitpoints less than your Max damage, you have a chance, dependent on their health, of wasting a roll higher than their health to kill them. If you compile all of your shots over a number of battles and don't drop kill shots, you'll find that your average damage per shot is lower than the gun's mean damage due to this effect reason.
As for Max damage rolls, there is one additional reason we see less of them than Min rolls: Low Health Tanks. When you hit a tank that has hitpoints less than your Max damage, you have a chance, dependent on their health, of wasting a roll higher than their health to kill them. If you compile all of your shots over a number of battles and don't drop kill shots, you'll find that your average damage per shot is lower than the gun's mean damage due to this effect reason.
Now, if we had a data set to help us determine the rule WG uses to truncate their random numbers, we could figure out just how badly off my assumptions are! I'll see about getting Xylenes to get some data on that...
As always, if you enjoyed this post, please help us in our data gathering efforts. It's simple, it's easy and you can find out how just by clicking here.
As always, if you enjoyed this post, please help us in our data gathering efforts. It's simple, it's easy and you can find out how just by clicking here.
Monday, February 4, 2013
Technical Difficulties
To those of you who keep checking this blog, thank you very much. I hope to have a new analysis up in the next couple of days, but I have been having some computer issues, so it's taking much long than expected. Thanks for your patience and support.
Thursday, January 31, 2013
Yea, but how important is it?
Over the last few posts, I've been providing you with a look at battles on Province, the crazy low-tier map that's so unbalanced it makes your head hurt. With that look, I gave you a list of tanks and a number that represented how they affect your win chances if you plugged your team composition into a fancy formula.
Today I'm going to take a step back and try to inform you something about their usefulness: how well do they let us predict the outcome of battles?
I do this for two reasons:
1. The only way to know if these numbers mean anything is to apply them to data and see if you get accurate predictions.
2. People who know their statistics keep harping on me to do some legit hypothesis testing/other statistical things.
So, to evaluate how good my regressions are, I'm going to slip off into a slightly different data set: All Tier 1 Battles. 1576 battles were included in the model generation, another 1506 in the validation set (these were separated randomly). Again, T1 Cunninghams were used as the baseline tank.
Once again we see an interesting distribution, ~51% of games are won by side 1, however we can attribute it to:
A. Effects of Province on the total distribution.
B. Random chance.
We know A from the last set of analyses, while we know B from the P-value of the intercept, at a 28.9% chance of getting that value purely by chance, we can't reject the null hypothesis of battles being, on average, 50/50. So, I'm going to run it again, forcing this value:
Hey look, I made a table! But I've stuck something in here that you probably haven't seen before, the Odds Ratio. It's really not hard to understand, it's just the amount that your base odds (in this case 0.5, or 50%) is multiplied by for each unit increase in that variable.
Count 'em up: Even LTraktors, Even FTs, Side 1 has +1 MK. I, Side 2 has +1 MS-1, Even T1s. So to predict the outcome is pretty simple:
From these you can clearly see that doubling the amount of data definitely increased the significance of each of the coefficients except for the LoLTraktor. Further the difference between these two models is pretty small, which is a sign of stability in the outcomes. The intercept shows that 52% of all battles are won by side 1, but now the significance level is such that we're mostly left with option A for an explanation. But now we get a whole new set of Odds Ratios:
Today I'm going to take a step back and try to inform you something about their usefulness: how well do they let us predict the outcome of battles?
I do this for two reasons:
1. The only way to know if these numbers mean anything is to apply them to data and see if you get accurate predictions.
2. People who know their statistics keep harping on me to do some legit hypothesis testing/other statistical things.
So, to evaluate how good my regressions are, I'm going to slip off into a slightly different data set: All Tier 1 Battles. 1576 battles were included in the model generation, another 1506 in the validation set (these were separated randomly). Again, T1 Cunninghams were used as the baseline tank.
Once again we see an interesting distribution, ~51% of games are won by side 1, however we can attribute it to:
A. Effects of Province on the total distribution.
B. Random chance.
We know A from the last set of analyses, while we know B from the P-value of the intercept, at a 28.9% chance of getting that value purely by chance, we can't reject the null hypothesis of battles being, on average, 50/50. So, I'm going to run it again, forcing this value:
This obviously isn't much of a difference, but it will give us a good starting point. In summary, from best to worst:
| Tank | Odds Ratio |
|---|---|
| Leichttraktor |
1.09
|
| MS-1 |
1.019
|
| T1 Cunningham |
1
|
| Renault NC-31 |
0.965
|
| Renault FT-17 |
0.915
|
| Vickers Med. Mk. 1 |
0.901
|
 |
| Example Battle: Somebody got Annihilated! |
Overall Odds Ratio = (1.09)0•(1.019)-1•(0.965)0•(0.915)0•(0.901)1= 0.882 or ~9:10 odds
And since P = (Odds)/(Odds + 1):
46.86% Win Chance
And since P = (Odds)/(Odds + 1):
46.86% Win Chance
Obviously this isn't the best of odds for our eventual winners, so lets see how well it does on other data.
If we were to just choose side 1 to win the entire time, we would get it right ~50% of the time, 51.4% in the case of this data. How well does the model predict it? Well, if we compare it to the data we used to generate the model, it gets 735 cases wrong (By wrong, I mean it gave side 1 a >50% chance of winning but side 2 won, or vice-versa). Remember that we had a total of 1576 battles, so our error rate with this model is 46.6%. In other word, this model only predicts 2% more battles correctly than you would if you laid your money all on one side, 3.4% than if battles are truly balanced. That may be good enough for Vegas odds makers, but not that great for us.
Applying it to the test set, the results aren't particularly good either. In fact, if we apply this data to my test set of 1506 battle, it only get 50.6% right! That's terrible! Scarcely better than flipping a coin! Surely there has to be some way we can make it better?
Well the first thing to do is combine the two data sets. Stick them together and what do you get?
Applying it to the test set, the results aren't particularly good either. In fact, if we apply this data to my test set of 1506 battle, it only get 50.6% right! That's terrible! Scarcely better than flipping a coin! Surely there has to be some way we can make it better?
Well the first thing to do is combine the two data sets. Stick them together and what do you get?
 |
| On this one I forced a 50% Win Rate |
 |
| But not on this one. |
From these you can clearly see that doubling the amount of data definitely increased the significance of each of the coefficients except for the LoLTraktor. Further the difference between these two models is pretty small, which is a sign of stability in the outcomes. The intercept shows that 52% of all battles are won by side 1, but now the significance level is such that we're mostly left with option A for an explanation. But now we get a whole new set of Odds Ratios:
| Tank | 50% Model | 52% Model |
|---|---|---|
| Leichttraktor |
1.0268
|
1.0239
|
| T1 Cunningham |
1
| 1 |
| MS-1 |
0.9516
| 0.9495 |
| Renault NC-31 |
0.9179
| 0.9181 |
| Renault FT-17 |
0.9106
| 0.9096 |
| Vickers Med. Mk. 1 |
0.8583
| 0.8563 |
Looking at it this way we can see a new shift in the data, suddenly the MS-1 drops below the T1. But we still need to do the important part, see how well the model predicts.
So, of the total 3082 tier 1 battle, the previous model, based on half the data, had only predicted 1603 right, a tepid 52%. With these new models, we should see some drastic improvement, right?
Behold! The mighty power of the 50% Model! 1550 Correct predictions! An astounding 50.3%! Wait, that's actually WORSE than having no predictors at all! It's scarcely even an improvement on our 50/50 win split assumption! That's terrible! Maybe the other model will do better...
Behold! The mighty power of the 50% Model! 1550 Correct predictions! An astounding 50.3%! Wait, that's actually WORSE than having no predictors at all! It's scarcely even an improvement on our 50/50 win split assumption! That's terrible! Maybe the other model will do better...
The power of the 52% Model will astonish you with 1666 correct predictions! 54.05% correct! That's right, 60 more correct predictions! 2% additional accuracy! The slivers of probability that gambling careers are made of. But still not a tool you'd want to bet the house on. Obviously there's more to these battles than tank type.
So basically, the moral of the story is this: At least down here, at tier 1, the differences in tanks are so small that building a prediction tool based soley on tank types isn't going to help you much. No tank is so overpowered that it overrides skill or luck in determining the outcome of the match. rather the differences in tanks is just as likely to be a bias in the players who favor them, which is something we'll have to control for at some point.
So basically, the moral of the story is this: At least down here, at tier 1, the differences in tanks are so small that building a prediction tool based soley on tank types isn't going to help you much. No tank is so overpowered that it overrides skill or luck in determining the outcome of the match. rather the differences in tanks is just as likely to be a bias in the players who favor them, which is something we'll have to control for at some point.
As always, if you enjoyed this post, please help us in our data gathering efforts. It's simple, it's easy and you can find out how just by clicking here.
Tuesday, January 29, 2013
And the Winner is...
At this point I've probably bored you enough with my forays into telling you which tiers and nations are better on one weee 'ittle map visited only by tier 3 and under tanks. But I have one last big reveal, the full listing. Trying to estimate just how good all 54 tier 3 or less tanks are on Province represents a lot of typing, after all, those names are long, but for you, I did it.
Now, to get you ready, I used as my baseline the T1 Cunningham, so positive values indicate it's better to have that tank than a Cunningham, while negative values indicate you're better off with the Cunningham.
Honestly, I rather surprised by just how many of these produce significant results. If you read the previous posts, you'll notice that most of these values fall in line with the findings from their regressions: Artillery is bad, TDs good, Premium Tanks good. The two best tanks to have on your team are German Premiums. Tier 1&2 British tanks are poor shadows of their counterparts.
The best:
PzKpfw II Ausf J: This one really surprises me, since looking through the data I can only find one instance of it appearing in a battle. But the data suggests (with a surprising level of significance) that having one on your side gives you a nearly 30% increase in win chance. Obviously either the tank or those who drive it are extremely good...
PzKpfw S35: Another tier 3 German premium, this one give you an extra ~25% chance of winning compared to having a Cunningham.
M22 Locust: American Premium this time, boasting a respectable +23% win chance.
FCM36 PaK40: The French Tier 3 TD seems almost perfect for this map, long view range and powerful gun. It's slowness is not a problem as there isn't much movement to do. This translates to a +22% win chance.
PzKpfw 38H735.f: The infamous Micro-Maus gets the distinction of being the only Tier 2 to make the top 5. Each one on your side give +21% chance of winning.
Obviously it's easy to find a pattern to this data. Premiums are good, Artillery Bad, the lower the tier the worse the tank. It's pretty simple to understand why this is, most people don't play the Tier 1-3 for more than a couple weeks, and even then, rarely with elited tanks and good crews. Premium tank players on the other hand tend to have more experience and are willing to pay, both gold and credits, for better crews and ammo. Further, premiums have no XP grind, so you're never going to find a premium tank fighting at less that full effect.
In an attempt to verify this theory, let us take a look at two premium tanks in particular, the T1E6, this year's gift tank and the M3 Light, the Lend-Lease M3 Stuart that players could get by finishing the tutorial. One peculiarity of the M3 Light is that when you get it, it comes with a free 100% crew, giving players with little experience a fully qualified crew much sooner than they would hav otherwise.
The T1E6, on the other hand, was gifted to everyone with a 50% crew. If you were like me, you immediately dismissed them to place your trained Chaffee crew members in it. But some people forget that it is a premium tank and can use tankers trained on other vehicles.
Comparing the M3 and the T1E6, it's fairly obvious that the M3 underperforms for it's tier, the fact that 50% crew tier 2 and a 100% crew tier 3 perform similarly is ample evidence of that. Comparing to other tier 3 premiums, the skill differential can be estimated at about 0.8, or 16% win chance. This shows that experience is not an irrelevant factor, but having a good crew is also important.
One thing you may have noticed is that the map bias we noted in the last two post is a bit higher in the full tank run. Also, we have some rather extreme biases toward high skill players in the form of premium tanks. How about we try to cut out that bias by removing the premium tanks.
Ok, I messed up a little. The T1E6 stayed in and I don't access to my regression currently to fix it. But, other than a few top tier tanks, almost every tank was close to 5% significance level. And all the low tier ones were highly significant. Plus the map bias has tweaked a bit down toward where we've had it previously.
One thing that should pop out immediately is that almost all of these tanks have negative coefficents, the M3 Stuart being the sole exception (With about 45 predictors, it's rather amazing that we only had one that truely deviated. Further, the significance level indicates just how random that value truely is, so I am going to ignore it and wait for more data). This tells me that it's not so much a question of who is better, but of who is less bad. Obvious offenders appear again, Artillery tends to be close to -0.9, some TDs are very good, others not quite so good.
The surprise winner among these tanks? The French D2 Medium, though it too is a low significance result. Beyond that, the Chi Ha and Luchs make good showings. The most surprising tank would be the classic LoLTracktor (Leichttraktor) which performs 0.11 points better than it's fellow tier ones, though this may be attributed to nostolgia driving by high skill players.
One could make a whole chart of which tanks are good, bad and balanced. But I will leave that for another time.
As always, if you enjoyed this post, please take a moment and download our Replay Upload Tool. It's simple and fast to use and each replay improves the quality of our data and the results we can show!
Now, to get you ready, I used as my baseline the T1 Cunningham, so positive values indicate it's better to have that tank than a Cunningham, while negative values indicate you're better off with the Cunningham.
Honestly, I rather surprised by just how many of these produce significant results. If you read the previous posts, you'll notice that most of these values fall in line with the findings from their regressions: Artillery is bad, TDs good, Premium Tanks good. The two best tanks to have on your team are German Premiums. Tier 1&2 British tanks are poor shadows of their counterparts.
The best:
PzKpfw II Ausf J: This one really surprises me, since looking through the data I can only find one instance of it appearing in a battle. But the data suggests (with a surprising level of significance) that having one on your side gives you a nearly 30% increase in win chance. Obviously either the tank or those who drive it are extremely good...
PzKpfw S35: Another tier 3 German premium, this one give you an extra ~25% chance of winning compared to having a Cunningham.
M22 Locust: American Premium this time, boasting a respectable +23% win chance.
FCM36 PaK40: The French Tier 3 TD seems almost perfect for this map, long view range and powerful gun. It's slowness is not a problem as there isn't much movement to do. This translates to a +22% win chance.
PzKpfw 38H735.f: The infamous Micro-Maus gets the distinction of being the only Tier 2 to make the top 5. Each one on your side give +21% chance of winning.
Obviously it's easy to find a pattern to this data. Premiums are good, Artillery Bad, the lower the tier the worse the tank. It's pretty simple to understand why this is, most people don't play the Tier 1-3 for more than a couple weeks, and even then, rarely with elited tanks and good crews. Premium tank players on the other hand tend to have more experience and are willing to pay, both gold and credits, for better crews and ammo. Further, premiums have no XP grind, so you're never going to find a premium tank fighting at less that full effect.
In an attempt to verify this theory, let us take a look at two premium tanks in particular, the T1E6, this year's gift tank and the M3 Light, the Lend-Lease M3 Stuart that players could get by finishing the tutorial. One peculiarity of the M3 Light is that when you get it, it comes with a free 100% crew, giving players with little experience a fully qualified crew much sooner than they would hav otherwise.
The T1E6, on the other hand, was gifted to everyone with a 50% crew. If you were like me, you immediately dismissed them to place your trained Chaffee crew members in it. But some people forget that it is a premium tank and can use tankers trained on other vehicles.
Comparing the M3 and the T1E6, it's fairly obvious that the M3 underperforms for it's tier, the fact that 50% crew tier 2 and a 100% crew tier 3 perform similarly is ample evidence of that. Comparing to other tier 3 premiums, the skill differential can be estimated at about 0.8, or 16% win chance. This shows that experience is not an irrelevant factor, but having a good crew is also important.
One thing you may have noticed is that the map bias we noted in the last two post is a bit higher in the full tank run. Also, we have some rather extreme biases toward high skill players in the form of premium tanks. How about we try to cut out that bias by removing the premium tanks.
Ok, I messed up a little. The T1E6 stayed in and I don't access to my regression currently to fix it. But, other than a few top tier tanks, almost every tank was close to 5% significance level. And all the low tier ones were highly significant. Plus the map bias has tweaked a bit down toward where we've had it previously.
One thing that should pop out immediately is that almost all of these tanks have negative coefficents, the M3 Stuart being the sole exception (With about 45 predictors, it's rather amazing that we only had one that truely deviated. Further, the significance level indicates just how random that value truely is, so I am going to ignore it and wait for more data). This tells me that it's not so much a question of who is better, but of who is less bad. Obvious offenders appear again, Artillery tends to be close to -0.9, some TDs are very good, others not quite so good.
The surprise winner among these tanks? The French D2 Medium, though it too is a low significance result. Beyond that, the Chi Ha and Luchs make good showings. The most surprising tank would be the classic LoLTracktor (Leichttraktor) which performs 0.11 points better than it's fellow tier ones, though this may be attributed to nostolgia driving by high skill players.
One could make a whole chart of which tanks are good, bad and balanced. But I will leave that for another time.
As always, if you enjoyed this post, please take a moment and download our Replay Upload Tool. It's simple and fast to use and each replay improves the quality of our data and the results we can show!
Saturday, January 26, 2013
That Stiff Upper Lip won't help you down here!
Since my full spreadsheet is still unfinished (give me a break, I've unlocked three tier 8s over the last three days!) and I still need to see just how much information we can develop from the data, I decided to go ahead and take a closer look at how the tanks of different nations perform on Province.
Before I show you the numbers, I want to give you a better understanding of what they mean. All of my numbers are generated using what's called Logistic Regression, which a fancy way of saying we can assign a value to the factors that effect your chances of winning. We get to choose the factors we think will be important, grab them in a data set and throw them at the little computer daemons in a package called R to work their statistical magic and give us these values, which we must then interpret.
When you go to perform a regression in R (or when you do any statistical regression really), you have to give it a function to regress. So far I have (and for the near term I will continue to) functions that are pretty simple: WinChanceSide1 = a + b*NetArtillery + c*NetTDs being the one from the previous post (you can see this at the top of the image). I had another variable to add there, NetTanks, that I didn't include because it would 'overdetermine' the system (NetTanks+NetArtillery+NetTDs=0, so if you know two, you can always find the third). This meant that Tanks were the baseline for the regression and you can figure out the win chances for each team from the values generated.
Now that that's all explained (and hopefully makes sense) let's take a look at the breakdown by nationality:
Interesting, apparently, at least on province, the British low tiers are REALLY underpowered. In fact, that's almost as bad as the artillery from the last post, but much, much more significant. In fact, down here it seems being anything but German is a Bad Idea(TM). Now, once again, I have to caution that these values are fairly insignificant, except for the British, and nationality is mostly a non-effect. In fact, looking at the American values, despite popping up with a slight negative value in this regression, based on the small size of the effect and the significance, I think it can be safely said that the American sub-tier 3 tanks are almost perfectly balanced. More data is probably necessary to identify if the Soviet and German biases are real or imagined (insert plug for people to send us replays with our Uploader here).
Now there's one last category I want to take a look at before we go ahead and regress on all of the tier 1-3 tanks. Premium tanks are generally intended to be "better than a stock tank, but worse than a same tier elited tank." So, down at these low tiers, versus other low tiers, on a very specific map, the results may surprise. After all, people who play low-tier premiums tend to be very experienced and have a greater likelihood of having a good crew. But let's take a look:
Holy unbalanced Batman! Here is a category of tank that gives a HUGE increase in win rates. Having just one extra on your team gives you a whole 5% better chance of winning. That's pretty damn imbalalanced if you ask me...
So now we have three good predictors for winning on province: What side you are on, how many more British tanks do you have and how many more premiums. So I'm going to make us a simple model for guessing how likely you are to win on province, using these three variables:
This gives us a pretty simple, but generally accurate guess of win likelihood from just those three questions:
Chance of Side 1 Winning = (e -(0.275+0.182*NetPremiums-0.08*NetBritish) +1)-1
Chance of Side 2 Winning = 1 - Chance of Side 1 Winning
 |
| In case you forgot which side is which |
Again, this makes no allowances for draws, but as we have demonstrated, almost all of the draws come out of side 2's win rate, so we can estimate it as:
Chance of Side 2 Winning = 1 - Chance of Side 1 Winning - 0.03
Next time I plan to perform a regression over the whole of tier 1-3 tanks, which I expect to have a lot of low significance results until more data is available. Then we can move on to bigger and better things...
Please, if you enjoyed this post and want to see more like them, be sure to help by contributing your replays! It's fast and simple, just see how here.
Subscribe to:
Posts (Atom)