Re: [Bug-gnubg] Rotating rolls in rollout

[nis]

--On 04  Oct 2002  13:44 +0200 Jim Segrave <address@hidden> wrote:

> On Fri 04 Oct 2002 (13:23 +0200), Nis Jorgensen wrote:

> > Perhaps something like this (pseudocode)
> >
> > RollUsed(TrialNumber,Rollnumber) =
> >         Sum for k = 0 to (Rollnumber - 1)
> >                 of
> >         ((Trialnumber div (36^k)) mod 36)
> >                 mod 36
>
> What happens to the results if you do something simple minded like:
>
> create an array of rolls with
> 1296 elements. Initialise with 1111 1112 1113 .. 6664 6665 6666
>
> Use the seed to randomise the array - for example, generate two
> random nos. i and j in the range  0..1295 and swap array[i] with
> array[j]. Repeat at least 648 times.

That is not a good randomization strategy. You should do

For i = 0 to 1295
 create random j between i and 1295
 swap array[i] and array[j]
Next

Then by a simple proof by induction, the array is as random as your random 
numbers.

This does not, however, provide the maximum stratification if the number of 
trials is not a multiple of 1296. Thus I would prefer to put a little more 
effort into creating the array of rolls. My best idea so far is to cycle 
the first roll through the same sequence of 36 rolls, and cycling the 
second roll through the same sequence, but 'sliding' the sequence before 
each sequence of 36. Something like (still pseudocode:

Create array36 containing 11,12,...,66
Randomize Rolls36 as described before
Create empty array of 1296

for i = 0 to 35
for j = 0 to 35
array[i*36+j] = (array36[j], array36[i+j mod 36])

> Then use this array sequentially (and circularly, dropping initial
> doubles when rolling out initial positions) to supply the first two
> rolls. This gives you a pseudo-random but repeatable set of roll
> pairs and the processor cost is not very high - filling the array and
> randomising it is relatively cheap, after that, during the rollouts
> you don't have to keep track of what you've used.

I am not sure about "dropping the initial doubles" during the rollout. I 
think rather we should make an array of 30*36 for that case.

--
Nis Jorgensen
Amsterdam