When to set seeds

[john-harrold]

Hello I was working on some simulations and I was using both set.seed and rxSetSeed set to the same value. I was calling it right before I called rxSolve(), but it didn't seem to take effect (repeated running of the script was giving me different results). However when I placed the seed setting commands before the rxode2 model function definition it seemed to work. Am I imagining this? Or does the seed need to be set before the model function definition to work?

[mattfidler]

There is nothing in the model code about random number generation, so I'm unsure what is happening here.

[mattfidler]

There are tests for the behavior in test-seeds.R in these cases, the set seed is right before the rxSolve() call.

I double checked how rxSolveSEXP is called from R/C and it:
Is called from the Rcpp export which means that the internal R seed management is called within the Rcpp macros. (That is they call GetRNGstate() on entering the functon and PutRNGstate() on exiting the macro). (Internal Rcpp code starts here: https://github.com/RcppCore/Rcpp/blob/1e82f1872eebe4fc813161ec0bac9bdfc53e1619/src/api.cpp#L56)
All R seed management works OK.

When rxSetSeed() isn't called, then it uses a random number from R to seed the simulation (which is not ideal).

Each solve uses independent seed based on the initial seed selected + the simulation id number. Afterwards it increments the initial seed by the number of simulation subjects.

The only issue I noticed reviewing the code is this is not used in dop853, it calls the rxode2 seed in whatever state it is in.

Were you using lsoda followed by dop853?

If so perhaps that is the issue.

A reproducible example would really help here.

[john-harrold]

I apologize. I'm not sure what I was doing wrong before but I moved it right after the model definition and it's working now. Sorry.

[mattfidler]

Don't worry about it 😄