How to compute the complexity of a function inside the custom objective function?

[unary-code]

This is a question more for the SymbolicRegression julia package than the PySR package.

I am using the PySR package, and I am using a custom objective function.

1. Inside the custom objective function, I want to get the complexity of whatever tree is passed into the custom objective function. How do I do this? Without recursively searching the left child and right child of the tree, which I feel would take too long.

2. Also, is there a way to force the model to only use up to 2 exp() (e^power) operators in the functions it generates?

[MilesCranmer]

1. You can do it with compute_complexity(tree, options)
2. Yes, see nested_constraints: https://astroautomata.com/PySR/api/#working-with-complexities. If you mean limiting the total number in the expression, you will have to do that manually in the loss function.

For example, if we assume that exp is the 3rd operator in the list of unary operators, this would be:

num_exps = count(node -> node.degree == 1 && node.op == 3, tree)

Then you can return some large number if num_exps > 2. You can also do it like this if this is more readable:

num_exps = 0
for node in tree
    if node.degree == 1 && node.op == 3
        num_exps += 1
    end
end

[unary-code]

Oh okay, thanks for your answer.

I noticed that the "compute_complexity" function is in one of the example codes on the SymbolicRegression jl documentation website (https://astroautomata.com/SymbolicRegression.jl/stable/examples/), but there is no documentation for it. Not a issue, because this function seems easy to understand.

"nested_constraints" is what I want to use now, but the counting of the total number of "exp"'s in the expression might be useful later on. Thanks for explaining!

[MilesCranmer]

Oops, good point. It does have a docstring: https://github.com/MilesCranmer/SymbolicRegression.jl/blob/0f47c6baf783b436ccecd5e635f692516c92d963/src/Complexity.jl#L10-L24

but it seems like it doesn't show up on the website due to not being included here: https://github.com/MilesCranmer/SymbolicRegression.jl/blob/master/docs/src/api.md