[Bug-gnubg] How fast can you cheat??

[Frank Berger]

> > What I am denoting is that the NNP "learns" the attributes of the
> > specific roll generator you use.  And does it probably better than a
> > human can.

This is absolute nonsens.
Why? quite easy. Any NN I'm aware of is presented the position to  
evaluate it.
Therefore it never sees the dice and can therefore not learn a pattern.


> This is interesting.  The Mersenne Twister is linear, so in  
> principle it
> is "easy" to learn how to predict it.  Whether a neural net will in  
> fact
> learn how to predict it depends on a lot of things, but it's entirely
> believable.
The Mersenne Twister isn't cryptographic strength but fulfills quite  
some statistical tests.
AFAIK as long as you don't need cryptogrphic strengt MT is regarded as  
the best choice.
Therefore to train a neural net to predict something that is nearly  
undistinguish from real random
is an absolute waste of time. Even with a simple linear congruence  
random number generator
it's a waste of time.





> In addition to the possibility of using physically random bits, one  
> could
> also try using a cryptographically secure random number generator.
>
> http://en.wikipedia.org/wiki/Cryptographically_secure_pseudorandom_number_generator
>
> It would be interesting to see if training the neural nets with  
> different
> random number generators produces noticeably different results.
No. It wouldn't be interesting and no it wouldn't show any effect.
Even if a RNG was heavily biased  throwin e.g. 25% 3's and even if the  
NN would see the dice the effect would even out over the large number  
of different positions presented the net during learning. And even if  
there was an effect on the 8th post point digit it would be completely  
overridden by the effect of the initial randomization of the NN

Analyzing the effect of rng on NN seems to me as useful as analyzing  
conspiracy theory....