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

[Michael Petch]

On 21/08/09 2:57 PM, "Frank Berger" <address@hidden> wrote:

> Removing doubles would probably have an influence, but for any
> practial purpose even a linear congruent
> RNG is good enough so an NN wont learn any pattern
>

Agreed, removing double would likely have an influence. I made the
suggestion as an example of how machine learning may see the game
differently based on a bad PRNG. But I whole heartedly agree that its
extremely unlikely that the bot would pick up on such a tine and
insignificant various as to be seen as anything other than noise.
 
> AFAIK MT is not strong enough for cryptography but these cryptography
> algo are very challenging. For any practical purpose it is good enough.

Mersenne Twister in itself passes most tests of randomness (Hardened tests
etc). But from a cryptographic perspective it is bad because of
reproducibility. It can be attacked mathematically in such a way that one
could reproduce the seed and then the data given enough data and time. As
well, how the seed was originally generated is often targeted if you know
the method by which It was generated you can attempt to reconstruct it.

> 
> All this implies that there is an exploitable pattern in the RNG. I
> would bet that an NN trained to predict MT would fail

We don't disagree on this either.

But I wouldn't put bets on the old c/c++ implementations of random() in the
old days. They were highly periodic and very untrustworthy as a good source
of random numbers. Could an NN trained to predict that fail? I'd say most
likely it would fail too (Although it would have a better chance than
soemthing like MT)