How computer loads out ammo...

[ww2steel]

If anyone cares, I have figured out how the computer loads out ammo. ...yes, I get bored on my days off.

I took 220 SU122, 10 for each month august '43 through the end. It seems that once HEAT is turned 'on' that it does NOT simply pick a random whole number between a max and 0- for this vehicle it is 0 - 8. It actually does use a sort of 'bell curve'. This is kind of surprising since it is way more difficult to program, but pretty cool that they bothered. Granted that's not a very big sample size, but notice how few vehicles had 5 rounds... kind of wierd.

# of rounds / # of vehicles with this # of rounds

0 10

1 22

2 29

3 37

4 38

5 14

6 41

7 20

8 9

Anyway, it's a relatively appropriately weighted bell curve, just my sample size must be too small.

Hope someone finds this interesting!

Mike

[Kingfish]

220 is far too small a sampling for such an important test. Most folks here expect data from at least 12,000 samples, and even then its considered an average test. To get up to the level of John Salt, Rexford or JasonC you'll need to run +30,000.

[ww2steel]

not gonna happen that would take about a year of nonstop work to test this

maybe I can do it a different way, just use all one date... yeah I'll do that. but you still ain't getting 30,000+.

Mike

[ww2steel]

Okay, I just did another one, with 1002 SU122.

The results: (# rounds/ # of veh so equipped)

0 80

1 72

2 132

3 124

4 149

5 93

6 163

7 92

8 97

Is it just my imagination, or are the odd numbers getting shafted? Must be because 122 comes in boxes of two, right? :confused: As for the graph of this chart... I'm more confused than before. I guess it only took 30 mins to do 750 more vehicles, but I just ain't doin no more.

Anyway, the only thing I'm sure about is that it's a max of 8, min of 0, and it averages just barely over 4 rounds carried per vehicle.

Anybody run test on this before?

Mike

[76mm]

Luv ya ww2steel, keep up the good work.

[ng cavscout]

thanks for the info ww2steel... ummm, mightn't you want to go on a date or something? Maybe take the wife out for dinner if you are married?

J/K, but seriously, step away from the keyboard man.

[ww2steel]

I have to do a trip to Kentucky or something wed - fri this week, I'll give it a break then.

Mike

[John Kettler]

ww2steel,

While I salute your inquisitiveness and diligence, if you're going to go to all this time and trouble,

would it not also be worthwhile and insight producing to compute the standard deviations for each of these extensive analyses you've conducted?

It's certainly useful to know the mean number of rounds likely to be provided, but wouldn't it be more useful to know that and the deviation about that mean?

Regards,

John Kettler

[ww2steel]

I did have excel calculate this, but I think I'm doing it wrong. I took the standard deviation of each of the three tests as shown:

#rnds Test# (# of vehicles)

1(250) 2(376) 3(376)

0 12 39 29

1 26 22 24

2 35 52 45

3 40 40 44

4 42 53 54

5 14 38 41

6 45 52 66

7 23 35 34

8 13 45 39

Okay, so for zero rounds I calculated the standard deviation to be 11.15. When I graphed even two standard deviations plus or minus the totals of these three tests I still could not fit either a straight line or a bell curve into the margin created.

One big, obvious problem is that test 1 was smaller than the others, so of course it will throw off my stddev.

I assume I am supposed to use stddev for a population?

It has been ten years since my statistics class and I would really appreciate some help, as everything I can find online is either way too basic or more often way too advanced.

Please help! :confused: I'm really confused on this stuff and know I could do much better work if I could remmeber this stuff!

Thanks,

Mike

[John Kettler]

ww2steel,

It's been quite awhile since I last did this stuff

(about 15 years), but you appear to have enough samples (way more than the minimum 5 I recall as needed from Probability and Statistics) to get useful data. Each batch of tests should have a separately calculated mean, and each batch of tests can then be used to compute the standard deviation from the mean.

Let's say that Batch 1 yields a mean HEAT load of

4 HEAT rounds, with a standard deviation of, say, 2. Thus, most of the time, the player will get 4, shading off to 2 on the low end and as high as 6 on the high end, but that's true only ~70% (don't remember exact %) of the time. Why? Because that's for only one standard deviation. As you keep increasing the number of standard deviations,

you get more variability but less and less often.

You're working farther and farther out from the center of the bell curve.

Thus, each Batch should produce a mean and an associated standard deviation, which can be plotted and compared. Since Batch 1 has 126 fewer trials than the other two, I would expect it to be a bit more variable. Memory's rusty, but I believe two standard deviations cover some 90+% of the total sample population.

This should help. See particularly the graph of population % covered as a function of standard deviations.

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

Regards,

John Kettler

[Tim Hughes]

You do the SD calculation *down* the columns, not across the rows.

The calculation (assuming I haven't cocked-up somewhere) gives

Series Mean SD

1(250) 3.96 2.2

2(376) 4.16 2.5

3(376) 4.29 2.4

Sum(1002) 4.16 2.4

Mean given to 2dp and SD to 1dp

You should treat the results and their interpretation carefully because (a) the number of trials is still quite low (although with 1000 trials we should have random error down in the 3% range), and ( the distribution clearly isn't normal (consistent dip at 5), so the meaning of SD is not what one might expect.

[ww2steel]

Okay, that's what I was doing wrong, and why it didn't make sense to use the rows.

Now, how do I use this standard deviation to determine how likely say, the dip at 5 is to be real, and not statistical noise?

I need to just go audit another statistics class...

Thanks,

Mike

[Tim Hughes]

Originally posted by ww2steel:

Now, how do I use this standard deviation to determine how likely say, the dip at 5 is to be real, and not statistical noise?

Looks real to me

There are a number of sophisticated tools statisticians (and others) use to determine if an observation is significant. I haven't been a practicing physicist for 21 years now so I've forgotten it all.

In any case, the data here isn't really useful for what you want. Are you even sure the distribution is normal and not binomial? Poisson? Completely random?

See (for example) Does this data come from a normal distribution? for more ideas on checking - there's also the Kolmogorov-Smirnov test if you want to go that far (try Google). But just stuffing the data into OpenOffice and drawing a chart gives me severe misgivings about the normality of the distribution - it just doesn't look like it. If it's not normal then most high-school statistics aren't valid and speculation on the importance of the SD is close to pointless.

[Bannon DC]

All this effort... you could have spent the time playtesting scenarios from TPG. But, some people just love stats.

[Tim Hughes]

I can do statistics at work with some plausability (I'm on GMT/BST). Running full screen CMAK might be pushing it a bit

[ww2steel]

I first thought it indicated a simple random function in the program, but after running a graph of my first test it appeared to be a very distinct bell curve with some anomaly at 5, now... I just have no idea.

Mike

[Joachim]

Not sure... but ammo loadout may be a function of two (or more) random numbers with a fixed sum. This could result in a weird distribution when checking only for one variable.

They might have a distribution for each round with given percentages.

First round: p1% for HE, p2% for AP, p3% for HEAT, p4% for T with p1%+p2%+p3%+p4% = 100% = 1

The probability for having n HEAT rounds is then given by some formula including p3% and (1-p3%).

IIRC for p2%=0 and p4%=0 this should result in a distribution with two peaks.

But I'm too tired and some Glenfiddich Havana Reserve accompanying a Hoyo de Monterrey Epicure No.1 took its toll on my brain.

Gruß

Joachim

[Tim Hughes]

That's a binomial distribution. You won't see two peaks. It'll look just like a normal distribution at first glance.

The problem here is that the distribution is too flat to be normal or binomial - at least as they are usually understood. There just aren't any low or high tails. And the dip at '5' is so clear.

I find it easier to believe the chance is something like 45% for 0, 1, 5, 7, or 8 rounds and 55% for 2, 3, 4, or 6 rounds with the probability split evenly within each group giving:

</font><blockquote>code:</font><hr /><pre style="font-size:x-small; font-family: monospace;"># shots probability

0 0.09

1 0.09

2 0.14

3 0.14

4 0.14

5 0.09

6 0.14

7 0.09

8 0.09</pre>

[Mike Summers]

So the bottom line is you don't get too many rounds with your standard SU-122...is that what all this statistical analysis says?

I ain't none too bright about all that fancy math stuff.

[ww2steel]

I was just curious about how the rounds were distributed. I figured it would either be a bell curve (most tanks would have close to an average number with a few outliers that have lots or none... or a simple random number, where a tank would be just as likely to have any permissible number of rounds.

Just seems odd to me that 2 tanks in the same platoon would have such high variation in ammo loadout. "Why did Tank 2 get 4 Pzgr40 and I didn't get any?"

Mike