Vensim-Python Random Number Generator

[aliakhavan89]

Is it possible to reproduce Vensim random number generation process in Python?

If I have only the following variable in a model (attached):

Code: Select all

RANDOM NORMAL( -1e+07 , 1e+07 , 0 , 0.5 , 41 )

And I use the following code:

Code: Select all

import numpy as np

np.random.seed(41)

output = []

for i in range(100):
    test = np.random.normal(0, 0.5)
    truncated_test = np.clip(test, -1e+7, 1e+7)
    output.append(truncated_test)

I don't get the same results:

Screenshot 2024-12-23 at 10.50.54 AM.png (1.13 MiB) Viewed 5376 times

As far as I know, both Vensim and Python use Mersenne Twister algorithm for random number generation. Could Vensim lower precision cause any differences?

[Administrator]

Is it possible to reproduce Vensim random number generation process in Python?

Probably not.

I don't get the same results:

This is not surprising.

As far as I know, both Vensim and Python use Mersenne Twister algorithm for random number generation. Could Vensim lower precision cause any differences?

Lower precision? Vensim works with double precision numbers, I suspect Python does as well.

You would need to set the starting seed in Python to the same one that is in Vensim. And even then, if numpy does not use the exact same code we do, it's unlikely you'll get the same results.

Vensim uses the SIMD implementation of Mersenne Twister,
https://www.vensim.com/documentation/le ... _simd.html

[tomfid]

Even if the code were identical, it might be hard to get them to match - they'd have to be using the same Mersenne prime internally, and the number of draws has to be identical (Vensim makes an extra draw during model initialization). Also, Vensim adds an arbitrary integer offset to the seed.

A more sensible test would be to draw a million numbers from each and use the Kolmogorov-Smirnov test (or similar) to verify that they have the expected distribution. We did this with the NIST test suite.

[aliakhavan89]

These are great details. Using the KS metrics makes sense. Thank you!