Re-using best equation strings

[KevinG1002]

Hello,

New user here (SymbolicRegression.jl) and I absolutely love it!

I was wondering if anyone knew a way to turn the equation strings that are obtained after n-iterations into a callable function so that, for example:

returned_equ = (((x₅ * -50) * x₁) + -2) could be easily turned into a callable like f = turn_to_callable(equation_str) with f(x_1, x_2, ..., x_n) = .... yielding the result of the computation in the equation string (a bit similar to what Polynomials.jl does). I am essentially looking for something that could be easily applied to any new vector [x_1, x_2, ..., x_n] without going through predict(mach, x_test).

Perhaps, to simplify things, how does predict(mach, x_test) work exactly?

[MilesCranmer]

Glad to hear you are liking it! :)

If you run report(mach) you will get a NamedTuple containing an element equations. You can use this to access the callable functions for each equation:

using MLJ, SymbolicRegression

# ...

fit!(mach)

equations = report(mach).equations

This is a vector of callable expression objects. You can call these with:

eq = equations[end]  # Most accurate expression is last one
# Note that the default selection is `report(mach).best_idx`

formatted_X = MLJ.matrix(X)'  # Convert to matrix, and transpose to column-major
options = SymbolicRegression.MLJInterfaceModule.get_options(model)  # Equivalent of `Options(...)`

eq(formatted_X, options)

This will evaluate the resultant equation on X.

Sorry that you need to call get_options manually right now which is admittedly not great. However, once MilesCranmer/SymbolicRegression.jl#326 finally merges, it will be much easier to use this – you won't even need to pass the options object anymore as it will be embedded in the expression.