I agree and disagree. The NN never sees the dice - agreed. However I
believe an NN is indirectly guided by the dice. If you took the
neural net
trainer and had Gnubg play itself again but this time set up the
random
number source to throw away all the doubles I am pretty sure how the
Bot
learns to play the game over time will change.
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
My opinion though is the Mersenne Twister is cryptographically
strong enough
and has such a massive period that the Bot would likely not be able
to see
anything (the period for standard 32 bit versions of mersenne are
2^219937 ?
1).
AFAIK MT is not strong enough for cryptography but these cryptography
algo are very challenging. For any practical purpose it is good enough.
If you trained with one random number generator and then play with
another
any thing the NN might have learned as a result doesn't matter. GnuBG
doesn't continue to learn while it plays humans (or have knwoledge
of the
generator being used during the actual match), so it can't possible
garner
knowledge from random number generators it may never have seen (or
were
implemented differently).
All this implies that there is an exploitable pattern in the RNG. I
would bet that an NN trained to predict MT would fail, therefore a net
that tries to learn BG should learn that?
_______________________________________________
Bug-gnubg mailing list
[email protected]
http://lists.gnu.org/mailman/listinfo/bug-gnubg
