How to constrain all the constants to be positive

[HouJayson]

Hi Miles,

I am a huge fan of PYSR. Currently, I am using PYSR to find mathematic models for human tissues, it works very well.

Small question: is it possible for us to constrain all the predicted constants to be positive? For example, in f(x)=aexp(bx)+cln(1-dx), a,b,c,d should be strictly positive.

Thanks

[MilesCranmer]

You could do this with a custom objective function, like:

function eval_loss(tree, dataset::Dataset{T,L}, options)::L where {T,L}
    # See https://astroautomata.com/SymbolicRegression.jl/dev/types/#DynamicExpressions.EquationModule.Node
    is_negative_constant(node) = node.degree == 0 && node.constant && node.val::T < 0
    # (The ::T part is not required, but it just speeds it up as then Julia knows it isn't `nothing`)

    # Will walk through tree and count number of times this function is true
    num_negative_constants = count(is_negative_constant, tree)
    #  (Tree utilities are defined here: https://github.com/SymbolicML/DynamicExpressions.jl/blob/master/src/base.jl,
    #  and let you treat an expression like a Julia collection)

    if num_negative_constants > 0
        # Return 1000 times the number of negative constants as a regularization penalty
        return L(1000 * num_negative_constants)
    end

    prediction, flag = eval_tree_array(tree, dataset.X, options)
    if !flag
        return L(Inf)
    end
    return sum((prediction .- dataset.y) .^ 2) / dataset.n
end

It's better to have a scaling penalty like this so that you give it a sense of direction (many negative constants are worse than one negative constant).

I am a huge fan of PYSR.

You made my day :)

Cheers,
Miles

[HouJayson]

Wow, amazing!! it works!! Thank you so much!

[frank-ccc]

Dear Miles and developers,
I am a huge fan of PYSR also, and I am very interested in loss development; how could I incorporate this loss or other custom losses in Python? not very clear to me. Many thanks

[MilesCranmer]

@frank-ccc could you post a new thread with the things you’ve tried and any issues? There are examples here: https://astroautomata.com/PySR/examples/. Also check through the discussions as there are now quite a few cool examples of custom losses :)