AnyLogic Beta Distribution is Wrong for Large P or Q

AnyLogic is a really old piece of simulation software, that like the manatee, exists because it has no real predators. Unfortunately companies still use it, so I've had to interact with it.

Probability distributions are math functions that can tell you what the probability of something happening is. One such distribution is the beta distribution. This distribution is a funky looking curve, defined between 0 and 1, which can look like any number of things depending on what the parameters α and β are (sometimes called p and q because mathematicians all have different favourite letters). Here are some examples:

AnyLogic has a whole bunch of different distributions built into it. They even have their own beta distribution, and a truncated variant! Wow!

Two working examples

To show the problem, let us compare the beta distribution in AnyLogic using the choose probability distribution wizard and SciPy, the scientific computing package in Python.

Using a Beta in the form Beta(α, β), let's first visualise a Beta(2, 5) in both pieces of software. The probability distribution wizard is a graphical interface in AnyLogic that allows you to plug in numbers for any available distribution, and it shows the output:

Look at that, the mean is a touch under 0.3. That's good, because according to the beta distribution page on Wikipedia I linked before, the mean of a beta distribution is actually always α / (α + β). This is ~0.28 for our case here, which looks about right.

This Python code should show the same thing:

import matplotlib.pyplot as plt
from scipy.stats import beta

# Parameters for the beta distribution
a = 2
b = 5

# Sample 100k points from the beta distribution
data = beta.rvs(a, b, size=100_000)

# Get value of the mean
mean = a / (a + b)

# Plotting the histogram of the samples
plt.hist(data, bins=100, density=True)

# Plotting the mean
plt.axvline(mean, color='black', linestyle='--')

plt.show()

And here is the output:

Take a moment to convince yourself these are the same. This is excellent! Now let's try some bigger numbers:

Here are the two plots side by side for these values. The mean should be ~5 because (5 * 1e9) / (5 + 1e9) ~= 5. (The 1e9 on the top comes from the scale. I need to do this otherwise AnyLogic's plot freaks out and doesn't show anything at all.)

They're the same!

A broken example

Ok, now let us try these parameters:

Ahh, 1e10, a number 10 times larger than 1e9. Now let's take a look at the distributions back-to-back:

Aaannnnddd we broke AnyLogic. How do we know it's AnyLogic that's broken and not Scipy? Well, because the analytical solution for the mean should be ~5. Scipy gets it bang on, and AnyLogic has curiously just returned the incorrect distribution.

Conclusion

I ran a couple more tests, and it seems that as p or q increase above int32_max, the samples begin to deviate from the correct distribution. I didn't investigate this much further than this, because with this knowledge I could solve my problem. But there we have it: AnyLogic Beta distributions break for p or q values larger than roughly int32_max.

I wish AnyLogic was open source so I could create an issue.

I wish AnyLogic threw an exception when p or q were too large.

I wish AnyLogic could put a warning about this in their woeful documentation.

I wish this team used SimPy instead.