On Apr 11, 2005, at 11:11 PM, Russell Standish wrote:
> I'm dealing with these questions in an artificial life system - Tierra
> to be precise. I have compared the original Tierra code, with one in
> which the random no. generator is replaced with a true random
> no. generator called HAVEGE, and another simulation in which the RNG
> is replaced with a cryptographically secure RNG called ISAAC. The
> results to date (and this _is_ work in progress) is that there is a
> distinct difference between the original Tierra PRNG, and the other
> two generators, but that there is little difference between HAVEGE and
> ISAAC. This seems to indicate that algorithmic randomness can be good
> enough to fool learning algorithms.
It definitely should be. At least certain types of cryptographic random
number generators are reducible to factoring. That means that if any
program can distinguish the output from the crypto RNG from the output
of a true RNG, you could factor a large number such as an RSA modulus.
This would be an important and completely unexpected cryptographic result.
Assuming that factoring is truly intractable, crypto RNGs are as good
as real ones, and deterministic universes are indistinguishable from