Use random number generators, not global seeds

[caleb-johnson]

Fixes #52

This PR removes the use of global seeds to control numpy randomness throughout the source

[caleb-johnson]

Actually the tutorial experiment no longer converges after these changes. Need to look further into it

[caleb-johnson]

Actually the tutorial experiment no longer converges after these changes. Need to look further into it

False alarm, the new generator was just giving a particularly bad sampling with that seed so config recovery couldn't bootstrap. Changing the seed brought things back to normal

[kevinsung]

The notebooks need to be updated too, to use a single random generator rather than repeatedly passing the same integer seed.

[caleb-johnson]

The notebooks need to be updated too, to use a single random generator rather than repeatedly passing the same integer seed.

So this means all of the functions need to be updated to accept a Generator? We can deprecate the usage of int for rand_seed and move to this. I just want to make sure I understand best practices

[kevinsung]

The notebooks need to be updated too, to use a single random generator rather than repeatedly passing the same integer seed.

So this means all of the functions need to be updated to accept a Generator? We can deprecate the usage of int for rand_seed and move to this. I just want to make sure I understand best practices

Any function that accepts a seed should immediately pass the seed to np.random.default_rng. So the type of seed is the type of a valid input of np.random.default_rng. I never quite figured out the correct type annotation for that, but I think int | np.random.Generator | None will do.

int is a valid type for the seed so there's no need to deprecate its usage.

[caleb-johnson]

int is a valid type for the seed so there's no need to deprecate its usage.

@kevinsung , thanks, I think I have it right now

[kevinsung]

This is an oddly specific seed, might as well keep using 42 or any other number. Also, the variable should be called rng for consistency.

[caleb-johnson]

The docs say to use a very large number. I figure this is better than 42 or writing out a bunch of digits

[caleb-johnson]

"
Seeds should be large positive integers. default_rng can take positive integers of any size. We recommend using very large, unique numbers to ensure that your seed is different from anyone else’s.
"

[caleb-johnson]

Updated the varnames in notebooks

[kevinsung]

The motivation for using a large number is to ensure that you don't accidentally choose the same seed as somebody else in the world. But this is not a concern here, where we only need our results to be statistically valid by themselves. Usually for these situations (testing, or just to make a notebook deterministic), you can use any integer, even 0.

But if you insist on using a very large number, then certainly 2**24 is not big enough. The doc you link suggests this:

secrets.randbits(128)  

But again, this is really unnecessary here.

[caleb-johnson]

It truly doesn't matter at all to me. I would think 42 (a number used by everyone in every context for everything) would be worse. I'll change it

[kevinsung]

Thanks. In this situation, 42 is fine. If we were going to publish scientific results, then we should indeed use a large number, such as 266529787016279213832423171828752297713.

[caleb-johnson]

OK I updated them all to 24 🤝