These forums have been archived and are now read-only.

The new forums are live and can be found at https://forums.eveonline.com/

Science & Industry

 
  • Topic is locked indefinitely.
Previous page123Next page
 

T2 Invention oddity

First post
Author
Previous Topic Next Topic
Steve Ronuken
Fuzzwork Enterprises
Vote Steve Ronuken for CSM
#21 - 2014-08-24 18:05:19 UTC
Falin Whalen wrote:
So what you are saying is, that you and Steve here can, with 100% accuracy, predict the outcome of each individual invention job based solely on the previous pattern of invention job successes/failures?

If so, then have you reported this bug/exploit to CCP?

If not, then you are deluding yourselves with apophenia.



I'm not saying that Smile

Woo! CSM XI!

Fuzzwork Enterprises

Twitter: @fuzzysteve on Twitter
Mister Falcons
NF Plains
#22 - 2014-08-24 18:22:41 UTC
Okay I done additional testing using 2 slots of invention at time and I have been constantly able to both produce expected of average of 46% results as well as constantly able to reproduce a much lower of 23% chance. I used total of 6 batches of 40 runs. I fully understand theory of random number generator.
Since for my sample I am working on Mining Crystals. I put lets use for example 2 scordite mining crystals BPC's. I have for example here made 2 600 runs copies. I start both tries for invention and now wait for results, I repeat for total of 40 tries and I end up with 17 T2 BPC. I retry with Omber BPC's this time. Again I have 19 T2 BPC. Lastly repeat with Velspare, and I have 20 success. Slightly in end for 120 tries I have average of 46.6667% win.
Now we try same but this time we use two different BPC's. For this test set I used Pyroxeres and Kernite each time. This time I only did 2 sets, to repeat 3rd time to expect a different result would be foolish. First set of 40 I got total of original list 9 BPC's. The Second attempt of 40 and we got 8 BPC's. Total for both attempts was 17 or 21.25%. I fully understand statical probability. I will do another test in 2 weeks when I have more skills and free Lab Slot available and see if same pattern exist for doing 4 BPC's at once.
Mister Falcons
NF Plains
#23 - 2014-08-24 18:42:24 UTC
Steve Ronuken wrote:
Of course, most PRNGs are seeded with data which may as well be random.

A common one is using data derived from the incoming data on an network interface. Sure, if you can record it, you could recreate it, but with it coming from so many different sources, any other way of recreating it is pretty much doomed to failure.

tl;dr:
Yes, you can predict it, if you know enough about the initial stated. No, you won't know enough about the initial state.


People who whinge about PRNGs not being truly random, for a purpose such as invention in Eve Online, are, hmm. I'm having trouble coming up with a phrase that the ISD wouldn't have problems with.


Sorry I got such a debate about PRNG's, that was not intent. I was attempting to illustrate an observation I made.
That was first attempt. I was later able by slight modification of approach to invention able to 1 reproduce a 46% average across 40 runs 3 times. I was also able to reproduce about a 22% average across two sets of 40 runs. Seems I have discovered a hidden mechanic that is was repeatable that can cause a reduced chance of success. Granted right now I have only tried this test with 2 Lab slots right now, but in 2 weeks I plan to retest again to see if we get same results when I test it with using same approach but this time with 4 slots.
Valixithor
Vengeful Legion
EMERGENCY BROADCAST
#24 - 2014-08-25 11:09:30 UTC
Done exactly 1000 invention runs this last week & came out with 462 BPC's. Chance of success was 48.xx% so I'd say it's there or thereabouts on target.

It is sickening when you deliver 30 jobs and end up with 6 BPC's, but it's also too easy to forget the times when you end up with 26/30.
Falin Whalen
Federal Navy Academy
Gallente Federation
#25 - 2014-08-25 13:47:57 UTC
Steve Ronuken wrote:
Of course, most PRNGs are seeded with data which may as well be random.



Steve Ronuken wrote:
Yes, you can predict it, if you know enough about the initial state. No, you won't know enough about the initial state.

So you can't predict it, and you only think you can see patterns. Not to say that patterns aren't there, you just can't predict them.



"it's only because of their stupidity that they're able to be so sure of themselves." The Trial - Franz Kafka 
Shadow Preldent
Deadspace Exploration Conglomerates
#26 - 2014-08-25 16:24:15 UTC
Steve Ronuken wrote:
Of course, most PRNGs are seeded with data which may as well be random.

A common one is using data derived from the incoming data on an network interface. Sure, if you can record it, you could recreate it, but with it coming from so many different sources, any other way of recreating it is pretty much doomed to failure.

tl;dr:
Yes, you can predict it, if you know enough about the initial stated. No, you won't know enough about the initial state.


People who whinge about PRNGs not being truly random, for a purpose such as invention in Eve Online, are, hmm. I'm having trouble coming up with a phrase that the ISD wouldn't have problems with.


Perhaps you members of CSM have access to the code base, but from my lowly state all i can do is look a result data.

I don't know how CCP generates the results of invention, but I can look at the success rate.

Based on how results compare to the expected results of a random system then we can concluded (with some level of certainty based on the size of the sample) that the output DOES or DOES NOT approximate a random. Any good PRNG would produce a result that statistically appears indistinguishable to random with sample sizes less than size of the series (2^16, 2^32, 2^64, whatever).

People like Falin Whalen who start off the conversation by saying the results are independent are making a couple mistakes.
1) The results are not independent of each other. Starting with the assumption that they are misses the entire conversation.
2) The results may or may not statistically appear to be independent, but you have to analyze results to figure that out.
3) That CCP is calculating invention in the most random way possible, because obviously there has never been bad design or bugs in any CCP code.

I'm not saying that invention results do/do not appear to be independent and evenly distributed. I am saying we should look at the data and make our judgement based on the actual results.

For example, if that data in the post was actually true . . .
adriaans wrote:
Track it over several thousand attempts and it is very close to what it should be.

For instance, for the past week I've had 14/100 on one character but 78/100 on another. Random is random, your sample sizes are WAY to small. Minimum for any sample size of value should at least be 1k+ attempts.


Assuming 50% probability of success then the probability of having a result outside the range 15->85 in a random system is 8.28e-14

Assuming 50% probability of success then the probability of having a result outside the range 13->77 in a random system is 1.19e-15

So a single person experienced two results in the same week each of which have a likelihood of less than one in a trillion?

We can make the following states from this data.
Either adriaans is lying about his data or we know with very high confidence that the system that produced the results is not random. Even thought the success rate of the combined runs is about right the data is way to "lumpy" to be produced by a random system (or in our case a system that does a GOOD job of appearing random).

-SP
Falin Whalen
Federal Navy Academy
Gallente Federation
#27 - 2014-08-26 01:16:05 UTC
Shadow Preldent wrote:
Even thought the success rate of the combined runs is about right the data is way to "lumpy" to be produced by a random system (or in our case a system that does a GOOD job of appearing random).

-SP

Real random data can be "lumpy" too. A game of roulette at the Monte Carlo Casino on August 18, 1913, when the ball fell in black 26 times in a row, or the longest recorded streak of one color in roulette in American casino history happened in 1943 when the color red won 32 consecutive times in a row, and one report of the number 19 landing seven times in a row.

It really doesn't matter anyway, if it is Independence, or PRNG results on rails, the law of large numbers makes it a moot argument anyway. I just wish people would stop whining about how their run of 30, only produced 6 T2 BPCs. Tough frigging luck buck o, sucks to be you right now. Either give me your stuff, and quit, or go back to inventing and keep your damned mouth shut about how, "CCP, MUST HAVE CHANGED INVENTION, OMG INVENTION IS RUINED, NO SUCCESS EVAR!!!1!1!!!1"

"it's only because of their stupidity that they're able to be so sure of themselves." The Trial - Franz Kafka 
Steve Ronuken
Fuzzwork Enterprises
Vote Steve Ronuken for CSM
#28 - 2014-08-26 11:47:39 UTC  |  Edited by: Steve Ronuken
The whole point about a random system, is that you can't predict what comes next, from what came before. Which means that yes, you'll get lumpy data.


(bear in mind, you get a bare fraction of the number of results. Because there are many other people doing invention as well)

Woo! CSM XI!

Fuzzwork Enterprises

Twitter: @fuzzysteve on Twitter
Shadow Preldent
Deadspace Exploration Conglomerates
#29 - 2014-08-26 17:51:12 UTC  |  Edited by: Shadow Preldent
Falin Whalen wrote:

Real random data can be "lumpy" too. A game of roulette at the Monte Carlo Casino on August 18, 1913, when the ball fell in black 26 times in a row, or the longest recorded streak of one color in roulette in American casino history happened in 1943 when the color red won 32 consecutive times in a row, and one report of the number 19 landing seven times in a row.


Well i'm not sure if you realize this because you are so bad at math but you are actually proving my point.
32 consecutive red results in an american roulette casino on 32 rolls has a probability of 4.13e-11
Which is approximately 500x and 20,000x more likely than results reported by one person in one week.
So either he was lying, there are a lot more people running invention jobs per week than have played roulette in the past century, or the results are not statistically random


Quote:
It really doesn't matter anyway, if it is Independence, or PRNG results on rails, the law of large numbers makes it a moot argument anyway. I just wish people would stop whining about how their run of 30, only produced 6 T2 BPCs. Tough frigging luck buck o, sucks to be you right now. Either give me your stuff, and quit, or go back to inventing and keep your damned mouth shut about how, "CCP, MUST HAVE CHANGED INVENTION, OMG INVENTION IS RUINED, NO SUCCESS EVAR!!!1!1!!!1"


Don't get mad bro, were just talking about math.
We know that results are not random (being generated by a computer and all). All we can ask if the results are distributed as expected by a random system.

Suppose I had two coin, one that is"fair" and one that is not. You pick up one of them and every time you flipped it the result came up tails. How many results before if was more likely that you picked the biased coin instead of the "fair" coin. How many results before you were 99% confident it was the "biased" coin. How many flips before you were 99.99999% confident it was the "biased" coin? You could never be "sure" it was the biased coin, but after enough flips you reach the point where scientists would agree that you had picked up the "biased" coin.

also, thank you for finally using your browser's built in spell checker.

Steve Ronuken wrote:
The whole point about a random system, is that you can't predict what comes next, from what came before. Which means that yes, you'll get lumpy data.


(bear in mind, you get a bare fraction of the number of results. Because there are many other people doing invention as well)


In a random system you can't predict what comes next and yes you will get lumpy data. Runs will happen, but we can say the probability of the results falling outside a certain range if the system was random. If the results fall outside a certain range an average more often than would be expected from a random system then you can say with a certain level on confidence (based on the results and the sample size) that the results were not generated by a random system.

I have never said one way or another if CCPs invention algorithm does a good job of approximating a RNG. I pointed out that if the data people present here is true (and I don't think it is), then the algorithm does not do a good job of approximating a random system. I also said that they only way to answer the question is with a large sample size of verified invention results (which I do not have) or by looking at CCP's code (ha!).


-SP
Netan MalDoran
Hail To The King
The Silent Syndicate
#30 - 2014-08-27 23:33:31 UTC
1: Don't waste decryptors on cheap items, they are usually only used with ships. (Still use cheap meta items though)
2: It is random, in my first week of inventing 100% of my jobs were successful, then in week 2, only about 20% of the were successful, now it's about 50%. What?

"Your security status has been lowered." - Hell yeah it was!

Falcon's truth
Falin Whalen
Federal Navy Academy
Gallente Federation
#31 - 2014-08-28 00:29:10 UTC
Netan MalDoran wrote:
1: Don't waste decryptors on cheap items, they are usually only used with ships. (Still use cheap meta items though)
2: It is random, in my first week of inventing 100% of my jobs were successful, then in week 2, only about 20% of the were successful, now it's about 50%. What?

Watch out, Shadow Preldent, will come back with "maths" to prove you are a liar, and that the odds of that happening are to great to be true, see it is all laid out in the spreadsheet he's made.

"it's only because of their stupidity that they're able to be so sure of themselves." The Trial - Franz Kafka 
Shadow Preldent
Deadspace Exploration Conglomerates
#32 - 2014-08-28 01:03:52 UTC
Falin Whalen wrote:

Watch out, Shadow Preldent, will come back with "maths" to prove you are a liar, and that the odds of that happening are to great to be true, see it is all laid out in the spreadsheet he's made.


Smart of you to stop using "maths" that you don't understand to try and prove your point. Or to paraphrase Sonny Curtis. "I fought the maths, and the maths won"

-SP
Owen Levanth
Sagittarius Unlimited Exploration
#33 - 2014-08-28 12:03:45 UTC
My last two inventions were both successful. According to my spreadsheet, this means a success-rate of 100%! This is far too much, CCP should nerf invention.
Falin Whalen
Federal Navy Academy
Gallente Federation
#34 - 2014-08-28 16:10:23 UTC
Shadow Preldent wrote:
Falin Whalen wrote:

Watch out, Shadow Preldent, will come back with "maths" to prove you are a liar, and that the odds of that happening are to great to be true, see it is all laid out in the spreadsheet he's made.


Smart of you to stop using "maths" that you don't understand to try and prove your point. Or to paraphrase Sonny Curtis. "I fought the maths, and the maths won"

-SP

Stupid of you for going, "That's impossible, the odds of that happening are astronomical. Here is the math to prove you wrong." When the events did happen.

Look I was like you, I had the Bayesian probability curves all worked out, set everything up for the lowest possible outcome even tho the odds were in my favor that that would not happen, and promptly lost almost everything due to bad runs going below even lowest expected returns for the amount of runs. The "maths" told me that the odds were virtually impossible for that to happen. The law of large numbers, apparently was taking some time out for a cup of tea, and wasn't returning my calls.

Bayesian statistics is just a mathematically pretty good guess as to the probability of certain outcomes, and the law of large numbers will eventually align the results to the mean. That still doesn't mean that the unlikely or statistically improbable can't happen.

"it's only because of their stupidity that they're able to be so sure of themselves." The Trial - Franz Kafka 
Ereshgikal
Wharf Crusaders
#35 - 2014-08-29 09:34:29 UTC  |  Edited by: Ereshgikal
Falin Whalen wrote:
Shadow Preldent wrote:
Falin Whalen wrote:

Watch out, Shadow Preldent, will come back with "maths" to prove you are a liar, and that the odds of that happening are to great to be true, see it is all laid out in the spreadsheet he's made.


Smart of you to stop using "maths" that you don't understand to try and prove your point. Or to paraphrase Sonny Curtis. "I fought the maths, and the maths won"

-SP

Stupid of you for going, "That's impossible, the odds of that happening are astronomical. Here is the math to prove you wrong." When the events did happen.

Look I was like you, I had the Bayesian probability curves all worked out, set everything up for the lowest possible outcome even tho the odds were in my favor that that would not happen, and promptly lost almost everything due to bad runs going below even lowest expected returns for the amount of runs. The "maths" told me that the odds were virtually impossible for that to happen. The law of large numbers, apparently was taking some time out for a cup of tea, and wasn't returning my calls.

Bayesian statistics is just a mathematically pretty good guess as to the probability of certain outcomes, and the law of large numbers will eventually align the results to the mean. That still doesn't mean that the unlikely or statistically improbable can't happen.


Of course you would be losing ISK since you based your investment on the assumption that it was a good PRNG behind it all. You just provided another good pseudo-sample supporting the view you are arguing against.

What others have argued in this thread is that the data (if true) points with fairly high degree of confidence to the PRNG not being a very good PRNG. It might be the case that it really is a very good PRNG, it is just that the data at hand shows that this scenario is very unlikely (if those calculations are correct).

What can we do? We could gather insane amounts of data and all surrounding parameters that we think could be used (network latency, time of day, length of characters last chat line, or whatever a paranoid schizophrenic could come up with) and then maybe would could find out how the PRNG works. Or, the easiest route is to get hold of CCPs code and analyze it...
Falin Whalen
Federal Navy Academy
Gallente Federation
#36 - 2014-08-29 20:07:18 UTC
Ereshgikal wrote:
Falin Whalen wrote:
Shadow Preldent wrote:
Falin Whalen wrote:

Watch out, Shadow Preldent, will come back with "maths" to prove you are a liar, and that the odds of that happening are to great to be true, see it is all laid out in the spreadsheet he's made.


Smart of you to stop using "maths" that you don't understand to try and prove your point. Or to paraphrase Sonny Curtis. "I fought the maths, and the maths won"

-SP

Stupid of you for going, "That's impossible, the odds of that happening are astronomical. Here is the math to prove you wrong." When the events did happen.

Look I was like you, I had the Bayesian probability curves all worked out, set everything up for the lowest possible outcome even tho the odds were in my favor that that would not happen, and promptly lost almost everything due to bad runs going below even lowest expected returns for the amount of runs. The "maths" told me that the odds were virtually impossible for that to happen. The law of large numbers, apparently was taking some time out for a cup of tea, and wasn't returning my calls.

Bayesian statistics is just a mathematically pretty good guess as to the probability of certain outcomes, and the law of large numbers will eventually align the results to the mean. That still doesn't mean that the unlikely or statistically improbable can't happen.


Of course you would be losing ISK since you based your investment on the assumption that it was a good PRNG behind it all. You just provided another good pseudo-sample supporting the view you are arguing against.

You don't have very good reading comprehension do you?

I based my assumption on that it was a BAD PRNG, and also calculated based on the worst probable outcome, with the law of large numbers and a healthy respect that it would eventually see me trough with enough trials. I still lost almost everything despite the many trials, I made enough profit off the succeses to cover datacores, and that was it.
Quote:
What others have argued in this thread is that the data (if true) points with fairly high degree of confidence to the PRNG not being a very good PRNG. It might be the case that it really is a very good PRNG, it is just that the data at hand shows that this scenario is very unlikely (if those calculations are correct).

What can we do? We could gather insane amounts of data and all surrounding parameters that we think could be used (network latency, time of day, length of characters last chat line, or whatever a paranoid schizophrenic could come up with) and then maybe would could find out how the PRNG works. Or, the easiest route is to get hold of CCPs code and analyze it


MIght as well read tea leaves, look at your horiscope, read the portens in the entrails of a goat, delve into all other forms of augury while youre at it.

"it's only because of their stupidity that they're able to be so sure of themselves." The Trial - Franz Kafka 
Ereshgikal
Wharf Crusaders
#37 - 2014-08-30 09:15:44 UTC
Falin Whalen wrote:
Ereshgikal wrote:
Falin Whalen wrote:

[---snip---]
Stupid of you for going, "That's impossible, the odds of that happening are astronomical. Here is the math to prove you wrong." When the events did happen.

Look I was like you, I had the Bayesian probability curves all worked out, set everything up for the lowest possible outcome even tho the odds were in my favor that that would not happen, and promptly lost almost everything due to bad runs going below even lowest expected returns for the amount of runs. The "maths" told me that the odds were virtually impossible for that to happen. The law of large numbers, apparently was taking some time out for a cup of tea, and wasn't returning my calls.

Bayesian statistics is just a mathematically pretty good guess as to the probability of certain outcomes, and the law of large numbers will eventually align the results to the mean. That still doesn't mean that the unlikely or statistically improbable can't happen.


Of course you would be losing ISK since you based your investment on the assumption that it was a good PRNG behind it all. You just provided another good pseudo-sample supporting the view you are arguing against.

You don't have very good reading comprehension do you?

I based my assumption on that it was a BAD PRNG, and also calculated based on the worst probable outcome, with the law of large numbers and a healthy respect that it would eventually see me trough with enough trials. I still lost almost everything despite the many trials, I made enough profit off the succeses to cover datacores, and that was it.


No, I did read your post and nowhere did you mention assuming a bad PRNG. But regardless of you assuming it was a bad PRNG or a good PRNG you must have assumed some sort of distribution which implies you must have based your assumption on already existing data. Otherwise "worst probable outcome" seems rather unfounded. It would be interesting (yes, really) to know how many runs you did and what your calculated "worst possible outcome" was; and the actual outcome. Data, man, data!

It would also be interesting to see the ~math~ applied to your data to see how unlikely your result really was given a perfect RNG.

Quote:

Quote:
What others have argued in this thread is that the data (if true) points with fairly high degree of confidence to the PRNG not being a very good PRNG. It might be the case that it really is a very good PRNG, it is just that the data at hand shows that this scenario is very unlikely (if those calculations are correct).

What can we do? We could gather insane amounts of data and all surrounding parameters that we think could be used (network latency, time of day, length of characters last chat line, or whatever a paranoid schizophrenic could come up with) and then maybe would could find out how the PRNG works. Or, the easiest route is to get hold of CCPs code and analyze it...


MIght as well read tea leaves, look at your horiscope, read the portens in the entrails of a goat, delve into all other forms of augury while youre at it.


Exactly. There is little point in trying to deduce the PRNG behind it all. I was being sarcastic about even knowing how the PRNG is seeded.
Myfanwy Heimdal
Heimdal Freight and Manufacture Inc
#38 - 2014-09-18 21:02:31 UTC
Simple problability in Real Life.

Let's say that you have something with 10% probability of success and you have a sample size of 5,000 events. What's the longest losing run you would expect in that run?

Most people wouldn't come close to the answer of the EXPECTED longest losing run of being 100.

The formula for working this out is: Log(RUNS) / -Log(PROB)

This effectively means that that the original poster isn't suffering anything untoward. But to be fair on him, this is very common misconception by people when it comes to probability and most people don't understand it. Or, rather, like the rules for Monopoly; they think that they know the stuff but don't.

Most people would fail to explain how the casinos in, say, Las Vegas in the US, manage to survive or even make a profit but knowing probability it's not too hard.

Anyway, I speak as someone who makes a living out of horse racing here in the United Kingdom and in such a field one really does need to know how probability works.

- Myfanwy

Pam:  I wonder what my name means in Welsh?Nessa: Why?
Patty Loveless
Sebiestor Tribe
Minmatar Republic
#39 - 2014-09-22 18:45:35 UTC
~1500 invention jobs this month. ~700 success.
Lumpy data is the norm i found. I went three sets (~30 jobs) getting 2 bpcs out of it, then got 22 prints out of 30.
Unless you are one of those paranoid people that think ccp specifically is targeting you, just keep inventing. Luckily, invention is pretty cheap and doesn't tie up too much isk. Everyone deals with the same chances.

The upcoming changes they have hinted to surrounding invention should help at least pad the losses side so people don't freak out so much (as no one ever complains when they are going 7+ / 10
Shadow Preldent
Deadspace Exploration Conglomerates
#40 - 2014-09-23 18:37:14 UTC
1) How is this thread not dead . . .

2) Too easy . . .
Myfanwy Heimdal wrote:
Simple problability in Real Life.

Let's say that you have something with 10% probability of success and you have a sample size of 5,000 events. What's the longest losing run you would expect in that run?

Most people wouldn't come close to the answer of the EXPECTED longest losing run of being 100.

The formula for working this out is: Log(RUNS) / -Log(PROB)


Of course if you evaluate those numbers, the result is ~80.8.

To get a longest expected run of 100, you would need a sample size of 40,000 events not 5000.

Can you please tell me exactly which booking agency you work for so I know where to make some very easy money?

Quote:

This effectively means that that the original poster isn't suffering anything untoward. But to be fair on him, this is very common misconception by people when it comes to probability and most people don't understand it. Or, rather, like the rules for Monopoly; they think that they know the stuff but don't.

Most people would fail to explain how the casinos in, say, Las Vegas in the US, manage to survive or even make a profit but knowing probability it's not too hard.

Anyway, I speak as someone who makes a living out of horse racing here in the United Kingdom and in such a field one really does need to know how probability works.

- Myfanwy


First you failed on your basic math and you then you tried to apply your failed math to the wrong question.

Nowhere in this thread is anyone talking about longest runs. We are talking about the probability of given outcomes from a sample size.

But yes, if you run 1 Million module Invention jobs with 40% chance of success (that will only take 20 years on a sinlge max invention char) you should expect to NOT see a run of 16 or more straight sucesses.
Previous page123Next page