RNGs, quality of solutions and repeat tests
- From: "Random Programmer" <nonexistent2032@xxxxxxxxxxx>
- Date: 3 Oct 2005 05:08:35 -0700
Hi guys. In my work I use random numbers a lot. All my runs start with
the same random seed (for consistency). Is this good practice? I am of
the belief that assuming your RNG is a good one, one random seed is as
good as any other (especially if you call the RNG many times). Of
course if your RNG is a bad one, the resulting sequence of random
numbers may have a significant bias depending on the seed.
Other researchers in the stochastic optimisation field sometimes
produce "averaged" results of optimisation runs with a selection of
random seeds, but seeing how my runs sometimes take a few days to
complete, obtaining a statistically significant sample is not
I use the "Park & Miller with Bays-Durham shuffle" method.
Any comments would be welcome.