| Pages: [1] 2 :: one page |
| Author |
Thread Statistics | Show CCP posts - 0 post(s) |
Luigi Vercotti
   |
Posted - 2011.05.14 04:29:00 -
[1]
I need to know if this is some serious aberration on my part, or something systemic in the game.
I am inventing T2 Mods. Have the racial skill and the datacore skills at level IV. From all that I read, I should have a 48.26% chance of success, since I use no decryptors.
Now, I have been keeping track of some small drone invention.
First block of 10x10 = 43/100.
So far in second block, 3/10;4/10;5/10;3/10 = 15/40.
That is now 58/140 = 41.42%, which is a far cry from 48.26%.
Is this something that others experience, or is my sample size not large enough yet?
|
Wendi Watson
Organization Too Secret To Know
|
Posted - 2011.05.14 05:13:00 -
[2]
 Originally by: Luigi Vercotti
my sample size not large enough
140 is a far cry from a decent sample size. |
Doctor Badger
|
Posted - 2011.05.14 06:44:00 -
[3]
On item X, rate of success should be 74.29%. Over the last 80 runs, I'm averaging a 82.38% rate of success. One 10-run install was 10/10 for success. Clearly invention is broken.
|
Dasola
Minmatar
|
Posted - 2011.05.14 08:32:00 -
[4]
Well you do have a 48.26% chance to be succesfull in any invention atempt. How ever it is probability, not garantie, so theres bound to be variations. Expesially in sample sizes that are small.
You could run 10.000 atempts to invent item X at probability 48.26% and still not get 4826 success. Becouse it depends randomness, you can get lucky and get 5321 success, or have a bad strike and get only 1283 success. in very large sample sizes, your success should how ever start to approach that 48.26%
When i was doing invention, i sometimes failed 10/10 trys, sometimes i was lucky and got 8/10.
|
Wyke Mossari
Gallente
|
Posted - 2011.05.14 09:19:00 -
[5]
Ignore the probability trolls, the success estimates are exactly that estimates.
There are at some empirical factors that do not feature in the guestimate commonly bandied around but do very much seem to affect actually the success rates.
1) What I call 'beginners luck', some kind of hard coded boost to the success rate on the first couple of dozen or so invention jobs. I've seen this on every one of my invention alts.
2) The reboot boost, whenever there is reboot there is a similar short term boost in success rate, I suspect this is triggering 1).
3) The all round researcher boost, characters that engage in other kinds of research, copying, ME/PE, feature an above the curve success rate.
|
Lutz Major
|
Posted - 2011.05.14 09:42:00 -
[6]
Originally by: Wyke Mossari
Ignore the probability trolls, the success estimates are exactly that estimates.
There are at some empirical factors that do not feature in the guestimate commonly bandied around but do very much seem to affect actually the success rates.
1) What I call 'beginners luck', some kind of hard coded boost to the success rate on the first couple of dozen or so invention jobs. I've seen this on every one of my invention alts.
2) The reboot boost, whenever there is reboot there is a similar short term boost in success rate, I suspect this is triggering 1).
3) The all round researcher boost, characters that engage in other kinds of research, copying, ME/PE, feature an above the curve success rate.
and of course you can proof this, because you made hundreds of invention alts to make the above not more than an assumption of yours.
By the way, probability is not an estimate nor guesswork!
|
FuzzyLogik360
Caldari
|
Posted - 2011.05.14 12:04:00 -
[7]
Originally by: Lutz Major
Originally by: Wyke Mossari
Ignore the probability trolls, the success estimates are exactly that estimates.
There are at some empirical factors that do not feature in the guestimate commonly bandied around but do very much seem to affect actually the success rates.
1) What I call 'beginners luck', some kind of hard coded boost to the success rate on the first couple of dozen or so invention jobs. I've seen this on every one of my invention alts.
2) The reboot boost, whenever there is reboot there is a similar short term boost in success rate, I suspect this is triggering 1).
3) The all round researcher boost, characters that engage in other kinds of research, copying, ME/PE, feature an above the curve success rate.
and of course you can proof this, because you made hundreds of invention alts to make the above not more than an assumption of yours.
By the way, probability is not an estimate nor guesswork!
These posts always crop up regularly, almost as common as "afk cloaker whine" or "I mine so my minerals are free whine".
Thee best way to clear up T2 invention confusion (regarding success rates) is if CCP simply published their process/algorithm for this. I have no doubt that they will use a common random number generator algorithm, with a weighting effect (to achieve calculated success rate [in the limit]). I can also be pretty sure that this WILL NOT BE PROPRIETARY CODE (especially if they just post the pseudo-code). So CCP, just publish the data/code and put this one to rest.
My 100% gurantee for this is that the law of large numbers is your friend. But like a few have said already, you need a sample size in the (hundreds of) thousands, not the hundreds.
|
Brock Nelson
Caldari
T2 Technologies Unlimited
SRS.
|
Posted - 2011.05.14 13:41:00 -
[8]
Even if CCP did publish how invention success is calculated, it still wouldn't stop the idiots from whining about it.
Investor Relation | BSAC SE Listing |
Dorian Wylde
|
Posted - 2011.05.14 14:55:00 -
[9]
Originally by: Luigi Vercotti
sample size not large enough
This is the answer to every question/concern/troll/whine/ragequit post in every game where a percentage is the topic of discussion.
|
Lutz Major
|
Posted - 2011.05.14 17:58:00 -
[10]
Originally by: FuzzyLogik360
Thee best way to clear up T2 invention confusion (regarding success rates) is if CCP simply published their process/algorithm for this.
propability := baseInventionChance *
(1 + 0.01 * encryptionSkillLevel) *
(1 + (datacoreSkillLevel1 + datacoreSkillLevel2) *
(0.1 / (5 - metaLevel)) *
decryptorProbabilityMultiplier
IF getRandomNumberBetweenZeroAndOneMethod() <= probability THEN
SUCCESS
ELSE
BUMMER :(
END IF
There I coded it for you.
Originally by: FuzzyLogik360
I have no doubt that they will use a common random number generator algorithm, with a weighting effect (to achieve calculated success rate [in the limit]).
WUUUUT?
Originally by: FuzzyLogik360
My 100% gurantee for this is that the law of large numbers is your friend.
You are a GENIUS! Not that statistics live from that theorem in any way ... no no
|
Johnny Enaluri
|
Posted - 2011.05.14 19:01:00 -
[11]
Edited by: Johnny Enaluri on 14/05/2011 19:03:13
Funny thing about invention.. On my main invention toon I was getting pretty **** poor run 1 day and I started some jobs with my copy/manufacture alt by mistake, not paying attention to the toon I was on when starting jobs.
Turned out the lower skilled alt, ended up with 100% success on the 6 runs it did meanwhile my main inventor ended up with 3 out of 9.
I was like wtf?
|
FuzzyLogik360
Caldari
|
Posted - 2011.05.14 22:00:00 -
[12]
Originally by: Lutz Major
Originally by: FuzzyLogik360
Thee best way to clear up T2 invention confusion (regarding success rates) is if CCP simply published their process/algorithm for this.
propability := baseInventionChance *
(1 + 0.01 * encryptionSkillLevel) *
(1 + (datacoreSkillLevel1 + datacoreSkillLevel2) *
(0.1 / (5 - metaLevel)) *
decryptorProbabilityMultiplier
IF getRandomNumberBetweenZeroAndOneMethod() <= probability THEN
SUCCESS
ELSE
BUMMER :(
END IF
There I coded it for you.
Originally by: FuzzyLogik360
I have no doubt that they will use a common random number generator algorithm, with a weighting effect (to achieve calculated success rate [in the limit]).
WUUUUT?
Originally by: FuzzyLogik360
My 100% gurantee for this is that the law of large numbers is your friend.
You are a GENIUS! Not that statistics live from that theorem in any way ... no no
getRandomNumberBetweenZeroAndOneMethod() ... er no, that's a function not an algorithm - try this. The rest of your "code" is all well and good for describng the calculation of the success rate, but I'm more interested in the random-number generator.
"Weighting effect" - fine, poor choice of word. In effect, what I meant to say was a random number is generated with a Gaussian distribution, with mean-value centred on the "success rate". Is that better?
Yes, I am a GENIUS. Thankyou. And the "law of large numbers" rules, ok. 
|
Shira Elan
|
Posted - 2011.05.15 00:12:00 -
[13]
Originally by: Luigi Vercotti
I need to know if this is some serious aberration on my part, or something systemic in the game.
I am inventing T2 Mods. Have the racial skill and the datacore skills at level IV. From all that I read, I should have a 48.26% chance of success, since I use no decryptors.
Now, I have been keeping track of some small drone invention.
First block of 10x10 = 43/100.
So far in second block, 3/10;4/10;5/10;3/10 = 15/40.
That is now 58/140 = 41.42%, which is a far cry from 48.26%.
Is this something that others experience, or is my sample size not large enough yet?
It actually looks relatively rare - 1.8% of people expect to get that same result, and only 5% of people will have results worse than you. 87% of people will get 60-80 successes on 140 trials.
As others have said LLN is your friend.
|
Lutz Major
|
Posted - 2011.05.15 04:17:00 -
[14]
Originally by: FuzzyLogik360
getRandomNumberBetweenZeroAndOneMethod() ... er no, that's a function not an algorithm - try this. The rest of your "code" is all well and good for describng the calculation of the success rate, but I'm more interested in the random-number generator.
"Weighting effect" - fine, poor choice of word. In effect, what I meant to say was a random number is generated with a Gaussian distribution, with mean-value centred on the "success rate". Is that better?
It is completely regardless how the actual implementation of the RNG works.
It is the whole point of a RNG is to supply random numbers NOT some remotely cluster-distributed numbers.
|
FuzzyLogik360
Caldari
|
Posted - 2011.05.15 18:02:00 -
[15]
Originally by: Lutz Major
Originally by: FuzzyLogik360
getRandomNumberBetweenZeroAndOneMethod() ... er no, that's a function not an algorithm - try this. The rest of your "code" is all well and good for describng the calculation of the success rate, but I'm more interested in the random-number generator.
"Weighting effect" - fine, poor choice of word. In effect, what I meant to say was a random number is generated with a Gaussian distribution, with mean-value centred on the "success rate". Is that better?
It is completely regardless how the actual implementation of the RNG works.
It is the whole point of a RNG is to supply random numbers NOT some remotely cluster-distributed numbers.
I think in my haste to post I've not made myself clear. So apologies.
I think the implemetation of the RNG does matter, because that is at the heart of the OPs original issue - Why does my actual "success rate" at invention not equal the calcuLated "success rate"?
The answer to this is that a random number generator will produce a spread of results, with mean value centred on the "calculated" success rate - this does not mean however that you will actually achieve that success rate in practice. Only over a very large sample (law of large numbers, i.e. when n tends to infinity) will the actual succes rate tend towards the calculated success rate.
Now, the choice of RNG matters because no algorithm will prodcue a truly random sequence. Many people claim that invention sucess rates are rigged or do not seem to follow random behaviour. If there is any truth in this at all (which I doubt), then it might show itself up in the choice of RNG algorithm (some RNG alogrithms are very poor).
Imagine for moment that CCP uses an RNG that "seeds" itself based on the 24 hour clock time. Lets assume a very poor implementation that seeds itself on the hour unit (rather than the current second or millisecond). If a player regularly logs in every day at the same time to start their invention jobs, they will inadvertantly tirgger the same seed in the RNG. This will lead to a "predictable" random number sequence - ergo, it will appear as if the invention success rate has been rigged.
I doubt the above actually happens, but if there is any suggestion whatsoever that invention is rigged or not truly random, my first place to look would be the RNG algorithm.
Now look what you made me do, I've gone all nerdy and serious.
|
Ikserak tai
|
Posted - 2011.05.15 23:09:00 -
[16]
To hades with the RNG, sample size, a rule of large numbers. My invention successes have sucked for 6 months, and nothing will change that.
The player who mistakenly used his lower-skilled alt and got a great success rate versus his main invention alt who got a lousy result, you've hit the nail on the head.
Why do invention alts get penalized for training up their skills? That's the real crock of dung here.
YOU'VE NEVER ROCKED 'TIL YOU'VE UNDOCKED. |
Akita T
Caldari Navy Volunteer Task Force
|
Posted - 2011.05.16 00:04:00 -
[17]
Originally by: Ikserak tai
Why do invention alts get penalized for training up their skills?
The answer is quite simple.
They allegedly aren't. 
And in case they really are, it's a bug which you should report.
Most likely however, just another statistical outlier, nothing more, nothing less.
_
Make ISK||Build||React||1k papercuts
|
Mara Rinn
|
Posted - 2011.05.16 03:46:00 -
[18]
I prefer my "chance of success" random number generators to provide linear, not Gaussian, distributions.
--
[Aussie players: join ANZAC channel] |
Athar Mu
Gallente
Federal Navy Academy
|
Posted - 2011.05.16 07:10:00 -
[19]
I occasionally (every few months) add up all my invention jobs and then my successes and work out the over all %. It's never exactly on that number (my sample sizes are usually 5000-10000) but its never far off, maybe a 1 or 2% either side.
Has anyone noticed that if you do a bunch of different modules there is always one module that sucks and one that tends to do well?
Just a note; recently I have been getting better results than usual, averaging around 55%-60%, I must say its been awesome, upped my profits a bit over the last couple of weeks. Its clearly broken because this has been going on for awhile (more than twice) and I hope CCP never fix it because that would blow.
|
Lutz Major
|
Posted - 2011.05.16 09:14:00 -
[20]
Originally by: Mara Rinn
I prefer my "chance of success" random number generators to provide linear, not Gaussian, distributions.
Exactly. The Gauss curve and RNGs work only together to display the knowledge distribution about said generators:
Originally by: FuzzyLogik360
The answer to this is that a random number generator will produce a spread of results, with mean value centred on the "calculated" success rate
You are really, really wrong in your assumptions! If the RNG provides any 'favorite' numbers, the R is not really random, isn't it?
|
RaTTuS
BIG
Gentlemen's Agreement
|
Posted - 2011.05.16 10:30:00 -
[21]
My response
|
FuzzyLogik360
Caldari
|
Posted - 2011.05.16 15:53:00 -
[22]
Originally by: Lutz Major
Originally by: Mara Rinn
I prefer my "chance of success" random number generators to provide linear, not Gaussian, distributions.
Exactly. The Gauss curve and RNGs work only together to display the knowledge distribution about said generators:
Originally by: FuzzyLogik360
The answer to this is that a random number generator will produce a spread of results, with mean value centred on the "calculated" success rate
You are really, really wrong in your assumptions! If the RNG provides any 'favorite' numbers, the R is not really random, isn't it?
I have used the term RNG perhaps to losely, when actually I should have used GRNG (or Gaussian Random Number Generator). I thought I had implied this, even indirectly, but perhaps not.
As correctly pointed out, a basic RNG favours no particular number and produces "white noise" or a flat profile. What I imply however is that CCP must use an implementation that fits this to a Gaussian probability distribution, which will be centred on the "calculated" success rate.
In fact, because the probability distribtuion (by definition) must be capped between 0.0 and 1.0, for very low "calculated" rates of success, the distribtuion is more likely to be Poisson (and the mirror image profile for high chance rates). But that's another discussion.
Either way, my original point still stands that there are lots of ways to mess up the implementation of the RNG. It would be instresting to see how CCP do it, even though I strongly suspect it ISN'T broken.
I don't believe that one or two people's anecdotal accounts of invention not delivering their expected profit margins is evidence that invention is malfunctioning. Again, the rule of large numbers... rules.
|
Nahkep Narmelion
Gallente
CALIMA COLLABORATIVE
Atrox Urbanis Respublique Abundatia
|
Posted - 2011.05.16 16:08:00 -
[23]
Originally by: Luigi Vercotti
....or is my sample size not large enough yet?
Actually your sample size is 14. You've run 14 invention jobs, not 140. So your sample is way to small, IMO. I would argue that you should to at least 86 more invention tries and then calculate your results vs. expected.
And keep in mind you are one of hundreds if not thousands of people doing invention. Some of them are just going to have crap luck, and you could be one of them. In fact, if there are thousands of them I'd say the chances are really good that one or more inventors will have crap luck at least for some subset of their invention runs.
If we can get more than say 25% of inventors posting that they are more than 3 standard deviations away from what they expect to get, then and only then would I start to worry that invention was bugged.
|
Adrian Idaho
|
Posted - 2011.05.16 16:41:00 -
[24]
Originally by: FuzzyLogik360
[What I imply however is that CCP must use an implementation that fits this to a Gaussian probability distribution, which will be centred on the "calculated" success rate.
Why?
@Lutz: that graphic is awesome! I'll steal your idea for future use 
|
FuzzyLogik360
Caldari
|
Posted - 2011.05.16 17:38:00 -
[25]
Originally by: Adrian Idaho
Originally by: FuzzyLogik360
What I imply however is that CCP must use an implementation that fits this to a Gaussian probability distribution, which will be centred on the "calculated" success rate.
Why?
Because random error exhibits a Gaussian or normal distribution. The invention process is a designed to look like a "trial-and-error" mechanism (from a game POV) with a weighted chance of success. This behaviour is best represented by a Gaussian probability distribtuion.
|
Adrian Idaho
|
Posted - 2011.05.16 19:01:00 -
[26]
Originally by: FuzzyLogik360
Originally by: Adrian Idaho
Originally by: FuzzyLogik360
What I imply however is that CCP must use an implementation that fits this to a Gaussian probability distribution, which will be centred on the "calculated" success rate.
Why?
Because random error exhibits a Gaussian or normal distribution. The invention process is a designed to look like a "trial-and-error" mechanism (from a game POV) with a weighted chance of success. This behaviour is best represented by a Gaussian probability distribtuion.
Let me rephrase that: what problem would this solve?
|
FuzzyLogik360
Caldari
|
Posted - 2011.05.16 19:29:00 -
[27]
Originally by: Adrian Idaho
Originally by: FuzzyLogik360
Originally by: Adrian Idaho
Originally by: FuzzyLogik360
What I imply however is that CCP must use an implementation that fits this to a Gaussian probability distribution, which will be centred on the "calculated" success rate.
Why?
Because random error exhibits a Gaussian or normal distribution. The invention process is a designed to look like a "trial-and-error" mechanism (from a game POV) with a weighted chance of success. This behaviour is best represented by a Gaussian probability distribtuion.
Let me rephrase that: what problem would this solve?
:P
The "problem" is that some people don't seem to understand why their "actual" rate-of-success at invention jobs does not equal that of the "calculated" rate-of-success using the following:
proability = baseInventionChance *
(1 + 0.01 * encryptionSkillLevel) *
(1 + (datacoreSkillLevel1 + datacoreSkillLevel2) *
(0.1 / (5 - metaLevel)) *
decryptorProbabilityMultiplier
The reason is that the chance of success of a single invention is only weighted against the above formula, but is otherwsie random. But...
If I was to perform 1000 invention jobs, and divide them up into batches of 10 I could monitor how many successes I get for every 10 jobs. If I then plot a frequency histogram of my results, low-and-behold, it would look like a Gaussian distribution almost* centred on a success-rate equal to that calculated with the formula above.
*almost - because you'll never actaully achieve the calcuated rate exactly (and this is the descrpancy that all EVE players will experience). The agreement between "actual" and "calculated" gets better the larger the sample however.
|
Greg Huff
|
Posted - 2011.05.16 19:32:00 -
[28]
Reading this is like watching 2 mice race to then end of an exercise wheel. 
|
Lutz Major
|
Posted - 2011.05.16 19:32:00 -
[29]
Originally by: FuzzyLogik360
I have used the term RNG perhaps to losely, when actually I should have used GRNG (or Gaussian Random Number Generator). I thought I had implied this, even indirectly, but perhaps not.
You did. But I think your assumption of CCP using a normal distributed RNG is still wrong
It's a simple coin-flip with a chance of X.
Consider a 50% chance (from 1 - 100): given normal distributed RNGs with 1000 numbers and a deviation of 1 the results 'could' be:
47 48 49 50 51 52 53
0 72 271 367 231 56 3
So you achieve 367 successes in 1000 tries - or a probability of 36.7%
For 99% of retries - and I hope you will concur- the number of successes will never be in around 50% otherwise the GRNG is not working probably :-)
Other the uniform distribution of 1000 numbers from 1 - 100, where n <= 50 is a success (as it is 50% chance):
Success: 498
Failures: 502
Give or take for any other retry.
As for the matter, the EVE RNG is not only used in invention, it is used when determining hit chance, module vs. resists, module vs. module, and so on and so on.
|
Lutz Major
|
Posted - 2011.05.16 19:38:00 -
[30]
Originally by: FuzzyLogik360
If I was to perform 1000 invention jobs, and divide them up into batches of 10 I could monitor how many successes I get for every 10 jobs. If I then plot a frequency histogram of my results, low-and-behold, it would look like a Gaussian distribution almost* centred on a success-rate equal to that calculated with the formula above.
*almost - because you'll never actaully achieve the calcuated rate exactly (and this is the descrpancy that all EVE players will experience). The agreement between "actual" and "calculated" gets better the larger the sample however.
As I've done about 20.000 inventions in the last two years and I downloaded the IndustryJob.xml on a daily basis - i give it a try and do the histogram :-)
But give me some time, will ya?
|
| |
|
| Pages: [1] 2 :: one page |
| First page | Previous page | Next page | Last page |