Revisiting random seeds

[wlandau]

@shikokuchuo, thanks to your work on secretbase, targets now has a robust solution for reproducible parallel-safe random numbers. I am starting this discussion to see how this could improve mirai and/or crew. Both currently use L'Ecuyer as a default, and according to ?RNGkind, it "exhibits 6 clear failures in each of the TestU01 Crush and the BigCrush suite (L'Ecuyer, 2007)". What are your preferences for mirai?

I am thinking about switching crew away from the rs argument of mirai::daemon() and instead having crew::crew_worker() call set.seed(kind = "default", seed = secretbase::sha3(x = list(launcher, worker, instance, nanonext::mclock()), bits = 32L, convert = NA)) before it calls mirai::daemon(). Would I need to do anything else to make sure mirai does not change the algorithm back to L'Ecuyer?

[shikokuchuo]

I'm mindful of ropensci/targets#1139 (comment) I think Henrik's recommendation is spot on conceptually - hence as mirai already implements L'Ecuyer streams, it seems useful to allow them.

As for setting the seed in crew, if you intend to do that in crew::crew_worker() before you call mirai::daemon() that would be fine. Just make sure you are not specifying 'rs' in that case as otherwise mirai::daemon() will assign that to .Random.seed at the start of the call.

As for what a default should be, I'll need further time to consider. Using secretbase would allow seeds to be set independently and move to MT as you point out.

[shikokuchuo]

Practically, I don't think it's any more cumbersome to call mirai::nextstream() rather than secretbase::sha3().

Cet par, you'd prefer MT over L'ecuyer due to those failures. However, in the parallel computing context, the overlapping RNG ranges are, if not a bigger concern, at least one that is more narrowly defined for the domain our software has been written for. Hence I think logically from a human psychology perspective it is this risk that we have to address foremost. Even without looking empirically at which is more severe.

Of course, the targets-type reproducibility requirement means that it's a no-brainer to use secretbase::sha3() there, but at least for mirai there isn't the assumption that everyone is going to need reproducibility.

Does that make sense?

[wlandau]

Yes, I think this is well-reasoned. That makes sense.