Need a statistician (or, the dicebot is broken!)

[GeorgeBates]

Recently been listening to cough voices hargent! who shall remain nogghh-herra! nameless who say the ahem random.org dicebot hack compensates for rolls at one extreme with rolls at the other.

[sorry, bad cold!]

This assertion needed to be tested, so I collected the results of 599 consecutive DRs over three sessions game sessions with some guy.


What's striking here is that the top of the bell curve appears to be flattened with a lower-than-expected number of sevens bracketed by higher quantities of sixes and eights. The shape becomes even more strange due to a paucity of fives and tens.
The suggestion being offered for this strangeness is that the bot is "compensating" for strings of results at one end of the curve with stretches of numbers at the other end. This argument points to a suppressed number of sevens as in indicator of this offsetting phenomenon.
Random.org states that it, "offers true random numbers to anyone on the Internet." A true random number would be independent of any results that came previously, would it not?
Perhaps someone who has studied more math than me can say if these results are extreme enough to be significant. More than 600 rolls may be needed to see if this is really a trend. I will keep counting. Comments welcome.

 

[uckelman]

The test you need involves computing the serial correlation coefficient, which is a measure of how predictable a roll is from the previous roll. You'd also need three orders of magnitude more rolls for a good sample.

 

[GeorgeBates]

The test you need involves computing the serial correlation coefficient, which is a measure of how predictable a roll is from the previous roll. You'd also need three orders of magnitude more rolls for a good sample.

Hundreds of thousands of DRs? Really?

 

[uckelman]

I wouldn't bother testing such a small sample from a roller which says it can give you 10000 rolls in one shot. You'll have much higher confidence for rejecting or accepting the null hypothesis with a larger sample. The last time I ran statistical tests on the roller Vassal uses, I was working on millions of rolls.

 

[Gordon]

 

[GeorgeBates]

I wouldn't bother testing such a small sample from a roller which says it can give you 10000 rolls in one shot. You'll have much higher confidence for rejecting or accepting the null hypothesis with a larger sample. The last time I ran statistical tests on the roller Vassal uses, I was working on millions of rolls.

Are you able to share a summary from any of those tests? Perhaps a link if you've reported this on GameSquad previously?

 

[zgrose]

Recently been listening to cough voices hargent! who shall remain nogghh-herra! nameless who say the ahem random.org dicebot hack compensates for rolls at one extreme with rolls at the other.

And yet they failed to provide any evidence of it... I bet I know who it was...

But more importantly, the dice code is short and open source. If dice-ers want to do something useful, point out where in the code this claimed bias originates from and we can change it.

 

[GeorgeBates]

And yet they failed to provide any evidence of it... I bet I know who it was...

But more importantly, the dice code is short and open source. If dice-ers want to do something useful, point out where in the code this claimed bias originates from and we can change it.

Yeah, I try to be a healthy skeptic and look at the evidence. That game sure felt strange, though. Maybe I'll continue to keep track at least until I get to 10,000 rolls. That won't be too many games at 600/crack.

 

[uckelman]

Are you able to share a summary from any of those tests? Perhaps a link if you've reported this on GameSquad previously?

See here.

 

[Sparafucil3]

http://imgur.com/bwFWMqQ

 

[GeorgeBates]

See here.

Thanks. Two quick remarks:

• The question from the VASSAL site concerned results from a single 10-sided die, which of course has a very different distribution from the results of two 6-sided dice
• The response, "We’ve never been presented with any statistically significant evidence that the die roller has a problem" runs into the 'absence of evidence is not evidence of absence' fallacy

Nevertheless, until there is more evidence, all that can be said is that in this case we have an unusual distribution of results from those 600 rolls.
Has anyone recorded the distribution over 10,000 rolls? 100,000?

 

[trailrunner]

That game sure felt strange, though.

I've had some games with physical dice that felt that way. Play enough games and some will be just wonky.

 

[uckelman]

This is because physical dice are mostly garbage.

 

[trailrunner]

The response, "We’ve never been presented with any statistically significant evidence that the die roller has a problem" runs into the 'absence of evidence is not evidence of absence' fallacy

That fallacy may not be true for statistics. If you do 1 million trials each containing 1 million 2d6 rolls, and one of those trials contains a distribution that isn't right, that doesn't prove that the dice roller is flawed. It's just part of the statistics.

 

[trailrunner]

This is because physical dice are mostly garbage.

Maybe, maybe not. But it also supports the idea that you can't draw conclusions based on a small sample size.

ETA: In the link you provided was this caveat:

"Real dice, with the exception of casino dice..."

I use precision dice. Not sure if they're up to the standard of casino dice or not, but my point about small sample sizes still stands.

 

[Sparafucil3]

Thanks. Two quick remarks:

• The question from the VASSAL site concerned results from a single 10-sided die, which of course has a very different distribution from the results of two 6-sided dice
• The response, "We’ve never been presented with any statistically significant evidence that the die roller has a problem" runs into the 'absence of evidence is not evidence of absence' fallacy

Nevertheless, until there is more evidence, all that can be said is that in this case we have an unusual distribution of results from those 600 rolls.
Has anyone recorded the distribution over 10,000 rolls? 100,000?

You have two options for dice rollers in VASL: random.org and the built in roller. There is a paper on the VASL website with a very in-depth analysis of the built in one. The one at random.org isn't even pseudo random. It is truly random, pulling background noise out of the atmosphere to get true randomness. It is used by gambling organizations for generating random sequences. If they are satisfied, I am more than satisfied. There is always the voodoo of a camera focused on your dice tower/cup to remove all of the dice rollers from the equation. Of course, then there are questions about your dice, about your cup, about the surface you're rolling on, ... -- jim

EDIT to add: Of course, the random.org's dice roller pulls 1000 dr on the call to random.org. These are then stored in memory on your PC and distributed as needed. So while the sequence is generated randomly, if you can peek into the sequence then there is nothing random at all really.

 

[uckelman]

• The question from the VASSAL site concerned results from a single 10-sided die, which of course has a very different distribution from the results of two 6-sided dice

The analysis I did was on 40MB of raw RNG output. You could use that to roll 2d6 55 million times or treat it as 330 million coin flips. It makes no difference which you do for statistical tests---if something is wrong, it will show up either way.

 

[von Marwitz]

Recently been listening to cough voices hargent! who shall remain nogghh-herra! nameless who say the ahem random.org dicebot hack compensates for rolls at one extreme with rolls at the other.

[sorry, bad cold!]

This assertion needed to be tested, so I collected the results of 599 consecutive DRs over three sessions game sessions with some guy.
View attachment 29925

What's striking here is that the top of the bell curve appears to be flattened with a lower-than-expected number of sevens bracketed by higher quantities of sixes and eights. The shape becomes even more strange due to a paucity of fives and tens.
The suggestion being offered for this strangeness is that the bot is "compensating" for strings of results at one end of the curve with stretches of numbers at the other end. This argument points to a suppressed number of sevens as in indicator of this offsetting phenomenon.
Random.org states that it, "offers true random numbers to anyone on the Internet." A true random number would be independent of any results that came previously, would it not?
Perhaps someone who has studied more math than me can say if these results are extreme enough to be significant. More than 600 rolls may be needed to see if this is really a trend. I will keep counting. Comments welcome.

Forget dice superstition.

VASL - Frequently Asked Questions

vasl.info

von Marwitz

 

[asloser]

I really do not understand why someone obsesses about this. It has been proved statistically over and over again that VASL dicebot and/or random.org are not a probelm.

You are better off studying tactics and reading the rules if you want to be successful in ASL. Dice play a role, but you usually can adapt to the unforeseen results. Adaptation to exteme results is part of being a skilled ASLer!

 

[DVexile]

Nevertheless, until there is more evidence, all that can be said is that in this case we have an unusual distribution of results from those 600 rolls.

Actually, no you can’t say that at all since that distribution isn’t “unusual” by any accepted statistical definition. In fact, it’s well within the range of “expected” for 600 rolls.