S

- Thread starter SSA415
- Start date

S

I would email the guy on the ASLML who wrote that long post on the dicebot for an answerSSA415 said:There's a topic going on the ML about Dicebot. I don't play VASL (yet!) but I am interested in testing my (real) dice. Someone once gave me a list of distributions based on 250-300 rolls per die, IIRC, that would let you know if a die was "fair" with like 95% accuracy. Anybody got anything like that?

However, I believe that there are tools you can buy that will measure and weigh your dice to tell you how random they are (i.e. their balance). You can also purchase balanced dice; most dice are unbalanced I believe. However, I would say that the number of variables involved in asl probably minimizes the effect of slightly unbalanced dice.

I am not sure how valuable your free time is, but mine is precious.There's a topic going on the ML about Dicebot. I don't play VASL (yet!) but I am interested in testing my (real) dice. Someone once gave me a list of distributions based on 250-300 rolls per die, IIRC, that would let you know if a die was "fair" with like 95% accuracy. Anybody got anything like that?

Why not just invest $15.00 in a set of precision dice from Alex Key rahter than spend a few hours rolling dice, writing results, then doing an anaylsis ?

:idea: :idea: :idea:

That and the precision dice are ugly. :wink:

S

Nat

Roll your dice several hundred times. Record each roll (each die, not just the total).

For real dice (e.g. not a software program) simply looking at the number of occurances for each dr should be sufficient.

It AMAZES me how much time / energy people put into worrying about their dice. :? I own a pair of "balanced" dice that I bought from Alex Key, but I rarely use them.

If you're worried about your dice, just make sure you and your opponent use the same dice.

Sam

You can use my dice when you pry them out of my cold, dead hand. Same goes for my dice glass.SamB said:If you're worried about your dice, just make sure you and your opponent use the same dice.

So what were the the results?SamB said:I'm the guy who posted the 14,000 DRs in the VASL dice bot.

I did a quick and dirty chi-squared on 12 sets of 60 throws of a single die. One of the sets failed the 95% confidence level, meaning if it were a real life casino, they're ripping you off. Edit: Oh, and two other sets were very close to failing.

Of course, 60 throws is a small set and chi-squared is tough to use at smaller numbers, but that's where the pseudo-random dice bots break down.

I bet over 14,000 DRs everything was fine and dandy. That's not the issue. The issue is the seed of the pseudo-random numbers and the cylical nature of the dice bots. VASL just uses the Java rnd() function which produces these. Over large sets of data, it is uniformly distributed.

I think ASLers get upset when the roll 5,5 then 5,6 then 6,6. Probabilities are very against rolling that in real life. For the dice bot, not unlikely.

Anyway, as long as VASL uses the base rnd() function, this will be a problem. Better seeding of the random number generator can help.

All this is IMO, of course.

Fine, just fine.So what were the the results?

One guy on the mailing list wrote:

> I pumped all these into my handy spreadsheet to get a count. Over 14K

> DR, the values are pretty close to the expected distribution.

>

> DR Expect Actual Difference

> 2 389 396 +7

> 3 778 806 +28

> 4 1167 1154 -13

> 5 1556 1526 -30

> 6 1944 1903 -41

> 7 2333 2353 +20

> 8 1944 1967 +23

> 9 1556 1588 +32

> 10 1167 1114 -53

> 11 778 817 +39

> 12 389 376 -13

And another wrote:

>>that's a chi square of 8.62. Over 10 degrees of freedom. No question

>>that this bot generates rolls _well_ within the expected distribution.

Now, personally, I don't care. Just play the game.

Sam "Its not the dice, its your tactics. And fate..." Belcher

If you want to do your own analysis, you can download the data (in an Excel spreadsheet) at http://dicetower.com/DiceBotTest.xls It is 3.4 meg, so I hope you have a fast connection.....

Sam

Again, I would expect the results after 14,000 are fine. Also, I don't look at the resultant of the DR; I believe it can confuse the issue. I'm not a very bright person, so I look at individual die throws.

So I took the first 60 DR's (2 sets of 1 die thrown 60 times) from your first 1000, and plugged it into my spreadsheet. Well, that first die (Red) looks pretty suspicious.

The expected count for each result is 10 (meaning for 60 throws, each should be seen 10 times in a perfectly random world). Here's what your results show:

Result, Count

1 , 10

2 , 17

3 , 10

4 , 4

5 , 12

6 , 7

Lots of 2's and very few 4's. Now, I realize it only 60 throws and there is a chance that this could be duplicated in the real world. That's why you need the chi-squared analysis. The result is 9.8 which is within 95% confidence (I would like to say barely). Had the 1's and 3's been a little different, it would have failed, but since both ended up perfectly at a count of 10, it helped lower their chi-square value.

Now, I don't mean to rant on and on about this, but I have experience with the rnd() in computer languages and believe it to be junk. It's not aimed at VASL (a fine product). Personally, I hate the dicebot because of the cyclical nature of its results which I believe to be devastating in the ASL game.

14000? I downloaded the xls file, but could got 13000 DR). The distribution of

the 60 dr you pulled out of that data worked out fine. I don't know if you read the ASLML,

but I'm the spb that ran a number of tests on a set of 85450 dr---and all of those

tests (chi-square test of DR distribution, serial test on dr, coupon collector's test on

dr, and birthday spacings test on dr) were passed by the data. I did note that the

phase space model that I constructed of the data indicates some dependancies between

successive dr---as would be suggested by the fact that the PRNG is a LCG with the

constants recommended by Lewis, Goodman, and Miller. It isn't clear that this

dependancy has any effect on the suitability of the dicebot for its intended purpose (that

is, to model die rolls in a game between honest players).

I'd be positively ebullient to hear any meaningful test that indicates that this is not

the case. Really. But suggesting that a couple dr look `suspicious' doesn't do it.

Dicebot flakiness is a regular topic among VASL players. I don't hear FtF people talking about how ALL of their dice are suspicious.

I would propose that someone do a Thuring-type experiment. You could send several skeptics two sample sets of 100 DR each. One would be from the dicebot (preferably from real game logs) and one from real dice. They would then try to guess which one the dicebot sample is. If the program really is sufficient for its purpose, you should not get a result significantly different from 50% accuracy. I strongly suspect that the skeptics will be able to pick the dicebot sample with a significant accuracy. If not, then perhaps the dicebot defenders have a point.

As for those that say none of this matters and "just play", I'll say that it's really no fun having a scenario you've been playing for weeks blow up due to statistically freaky dice behavior. Yes, this will happen once in a blue moon with dice. It seems to happen quite often in my VASL games.

Disclaimer : VASL is a great thing. The dicebot sucks and VASL would be much better with improvements to it, but I'm still greatful that RK made such a large contribution to the community. I'd guess several of us wouldn't be playing if it weren't for VASL. Thanks, Rodney.

Scott McFarlane

This is totally bogus. Really.That's all fine and well, but as ASL/VASL players we all have plenty experience with real dice and with the Dicebot. Of course there's many arguments about perceptions being unreliable, etc., but the fact is that too many people (myself included) hold the opinion that the Dicebot has problems for there not to be some issues. I know people who won't even use the Dicebot for this very reason.

Analysis of DRs generated by the dice bot indicate that there is NO PROBLEM.

Perception is NOT reality.

Sam "Test" Belcher

I don't why I'm responding because no one is reading what I'm saying.

I did a test myself (not with Sam's data) with 12 sets of 60 throws. One set failed the 95% confidence level, with two others failing about the 97% level. I took the first 60 of Sam's 13,000+ throws and it almost failed the 95% confidence level. To me, this is an analysis that warrants further exploration.

I contend the problem is with the rnd() function in small sets of data. In other words, each throw of a die is NOT 1/6. Given a result, the following result is influenced by the previous one. This is inherent in the rnd() function.

I totally agree that in sets of thousands of throws the data follows a normal distribution. In that respect, the rnd() function excels.

Next, I believe this dependency to be important in the game of ASL. Low rolls are almost a success while high rolls are rarely wanted. ASL is not a game of craps.

As an aside, it is my philosophical belief that Perception IS Reality. I learned that after working years in risk assessment engineering. Never underestimate the power of the human mind. But I'm not here to argue that.

In fact, I'm not really arguing anything. I thought I'd share my analysis and opinion, but if people want to hit the DR button 25,000 times and say, "Look, it works", fine by me.

My (Canadian) advancing fire phase, lots of lame shots that have to be made. German SAN =4

First shot

2,2

SAN dr 6

Second shot

2,2

SAN dr 6

Third shot

2,2

SAN dr 6

Forth shot

2,2

SAN dr 6

I stopped shooting after that!

Furax

I contend the problem is with the rnd() function in small sets of data. In other words, each throw of a die is NOT 1/6. Given a result, the following result is influenced by the previous one. This is inherent in the rnd() function.

I also have been a victim of this typical behavior in VASL3/4.

Is this bogus? How so? If there is such high confidence that the dicebot is random, then there should be no concerns as to whether or not people can differentiate the dicebot from real dice. If the dicebot is OK, then no one should be able to tell the difference, or....?

Perhaps the analysis you reference is bogus. Using 14,000 DR! Never played a game with that many rolls and who cares what the distribution looks like anyway.

1 1 1 2 2 2 3 3 3 4 4 4 5 5 5 6 6 6. Hey, a perfect distribution!

As has been pointed out, good agreement with the expected distribution is trivial with large samples. What's more important for the dicebot's intended use is whether it has an unusually high tendency to go on short-term streaks. Like 4 boxcars in a row. That should happen maybe once in lifetime of ASL playing. The odds are literally a million to one. It's happened to me twice on VASL. Anybody else see this sort of thing too often?

Scott McFarlane

nevermind..

mispost

mispost