Hacker News Reader

I also don't understand, i solved (what i think is) the same problem by having a master rng whose output seeded individual rng, one for map generation, one for enemy ai, one for loot, etc...

View on HN

This depends on what RNG you are using. If the master and the secondary RNGs use the same algorithm without extra internal states (e.g. splitmix or LCG), you will be generating the same stream over and over with small shifts. The end result may not be obvious in your use case, but can be a catastrophe for statistical simulation or the implementation of dropout. PCG32 is a proper solution. Xorshiro provides another solution by using one simple function to advance the internal state by 2^64 iterations.

BTW, xoroshiro should not be used. See the last section of this:

https://lemire.me/blog/2017/09/08/the-xorshift128-random-num...

I was aware of that blog post as well as the counter argument by the xoroshiro author against PCG. I am not studying RNG and can't say which is more correct. To me, I use xoroshiro mainly because PCG64 is complex to implement. My applications do not need extreme randomness so far.

Reply on news.ycombinator.com