Reproducibility of designs from old versions about lhs

[bertcarnell]

Ulrike,

It's good to hear from you. You might have noticed that I added an orthogonal LHS to the package in this latest round. It's something you and I discussed many years ago. I'm certainly open to feedback if you have it.

I was aware of the random number generation changes, but thank you for the reminder anyway. I have had other developers ask me about the reproducibility of the lhs code in the past. Because I made a large number of changes, I'm reluctant to make random number reproducibility promises yet. If users need reproducibility for testing today, I suggest they save the design.

I feel like the maximinLHS algorithm is the weakest algorithm in my package. In particular, I think that it doesn't perform with the guarantee that most people expect from it. For example, I can't claim that maximinLHS(6,2) actually produces the uniformly best sample that maximizes the minimum distance. The first algorithm I wrote builds an lhs one point at a time, adding the best next point to the design among a group of candidates. The algorithm does not guarantee maximin. In my latest package iteration, I've added other algorithms to maximinLHS to get closer to maximin if users want to spend the computation time. I expect to move that code from R to C++ in my next improvement. Moving to C++ makes it hard to guarantee the same random draws again.

When I released this last version, I made sure to perform the reverse dependency checks for your packages, but I didn't notice any problems. I hope I didn't cause trouble for DOE.wrapper.

Rob

[ugroempi]

Thanks, Rob! So I take it that using a seed gives the same result short term under the same package version (and RNG setting), but consistency between package versions cannot be expected for the time being. Is that correct? Then I'll document this in DoE.wrapper.

Your changes didn't give trouble for my packages, only the change in RNG did ...

I like your new oa_to_oalhs functionality, and it has given me some ideas regarding the use of regular or non-regular arrays. I'll e-mail you about those outside of github.

Best, Ulrike

[bertcarnell]

Yes. Consistency between versions cannot yet be guaranteed. MaximinLHS is the one function most in danger of change. randomLHS will likely be stable.