New algorithm for random number generation

[Namir]

I have been tinkering with a set of new algorithms for uniformly-distributed random number generation (RNG). I test each algorithm using several long runs. I test various statistics related to each run of 10000 numbers (such as auto-correlation and various histograms). I end up calculator an empirical "goodness" factor based on these statistics.

Recently I bumped into an interesting algorithm while working with a family of RNGs that use ratios of random numbers generated using simpler/basic methods. I noticed the following:

Code:


1. Generate r1 (between 0 and 1) using basic RNG algorithm #1
2. Generate r2 (between 0 and 1) using basic RNG algorithm #2
3. Calculate R -= r1/r2
4. If R > 1 then R = 1/R
5. If R = 1 then R = 0
6. Return R as a random number

I noticed that the above algorithm often generates better random number than the underlying basic RNG algorithms.

I will publish (on my web site) my tinkering in a few week.

Namir

PS: You have some flexibility in choosing the basic algorithms in steps 1 and 2.