I was recently working on a simulator that uses pseudorandom numbers to introduce variance into the system. The seed was based on the current timestamp (and therefore changed with every run).
After checking in a new command-line flag to allow specification of a fixed RNG seed, a colleague asked me "Why does this matter?"
I didn't know where to start ... the advantages seem obvious:
- Ensures reproducibility between runs
- Reduces noise between A/B experiments
- Eases debugging
Instead of giving a lecture, I'd like to point to some existing article that elaborates on these points. Does anyone know a good article or text?
Aucun commentaire:
Enregistrer un commentaire