Re:Specific Problems with Other RNGs
by cpeterso on Wednesday November 11, 2015 @01:14PM
(#10548132)
Has there been any research into generating a PRNG using a genetic algorithm whose fitness function is the Crush and Diehard test results?
[ Parent
| on HN ]
Not quite that, but here's someone's approximation of this approach for an extremely fast hash:
So this was a fun hash that I came up with when I was looking
for a 128-bit hash for falkervisor's code coverage database.
This hash passes all of smhasher's tests, faster than any other
hash that does the same.
I made this up by randomly testing things until it passed
all of smhashers tests. It is likely that this hash is heavily
flawed and smhasher just doesn't stress its flaws.
https://github.com/gamozolabs/falkhashZilong Tan's fast-hash contains a generator which
genetically optimizes over a combination of the cheap ops mul, xorshift left, xorshift right and rotate right, and also does genetic mutations over the initial magic states.
https://code.google.com/p/fast-hash/source/browse/trunk/hash...
The fit function is only the avalanche test, but that's easily exchangable.