Is there a way to prevent certain combinations of operators, like exp(cos(x)) ?

[jburz2001]

Hello. I'd like to use this package for Symbolic Regression of physics equations.

Now, physics equations rarely exhibit exponentials of sinusoids. So is there a way for me to tell PySR to not use exp(cos(x)) ? I know that I can constrain the complexity of exp to 1 through constraints = {'exp':1}, and that this would remove sinusoids, but I'd like to prevent exp from having the cos function as an argument altogether.

Can this be done? If so, how may I do it? I was thinking about defining my own exponential function that cannot take sinusoids as an argument, but I am currently unsure about how to do this.

Thank you!

[MilesCranmer]

Hi @jburz2001,

Great question, this is something that I think would be useful for many. You are right that you could disable complex arguments with the constraints option, but unfortunately there is no current way to disallow specific patterns in the equation like you are interested in.

While I don't have a timeline for when I would be able to implement this, what I can do is point you to where this could potentially be added if you are interested.

In the backend (https://github.com/MilesCranmer/SymbolicRegression.jl/), there is a function here:

function check_constraints(tree::Node, options::Options, maxsize::Int)::Bool
    if countNodes(tree) > maxsize
        return false
    end
    for i=1:options.nbin
        if options.bin_constraints[i] == (-1, -1)
            continue
        elseif flagBinOperatorComplexity(tree, Val(i), options)
            return false
        end
    end
    for i=1:options.nuna
        if options.una_constraints[i] == -1
            continue
        elseif flagUnaOperatorComplexity(tree, Val(i), options)
            return false
        end
    end
    return true
end

This checks the user-passed constraints, and returns false if they are violated. It's used during the equation search, so you shouldn't need to change anything else. Here is where you could try to set up a more complex constraint check.

If interested in trying to implement this:

1. An equation (I call equation variables tree, and the equation type Node) is a simple binary tree, defined here. You would want to recursively traverse it to test your constraint. You can see some example functions that operate on equations here.
2. You can work directly with the Julia backend independently (see the readme), or if you want to test it from Python, you can do: julia_project="[path/to/SymbolicRegression.jl/repo]" as an argument to PySR, and it will use the local Julia code instead of the package.

Let me know if you have questions!
Cheers,
Miles

[jburz2001]

Thank you for your detailed response, Miles. Apologies for the delay in mine.

I will look into your suggested methods for disallowing specific operator combinations. Thanks again!

[MilesCranmer]

Just following up on this. There is now the nested_constraints parameter, which is exactly built for this purpose.