It is impossible to determine more than 4 coefficients in the function

[CleverCompiler-oss]

Good day!
I am trying to predict 5 or more coefficients in a polynomial.
For degrees less than 5, this works. But starting with 5 different parameters, an error occurs:
"JuliaError: TaskFailedException
nested task error: BoundsError: attempt to access NTuple{6, Symbol} at index [7]"

I am using the following code:

import numpy as np
import matplotlib.pyplot as plt
from pysr import PySRRegressor, TemplateExpressionSpec
X = np.random.randint(1, 21, 1000)
Y = np.array(1.2*X + 3.1*X**2 + 6.1*X**3 - 11.7*X**4 -0.9*X**5 + 47)

template = TemplateExpressionSpec(
    function_symbols=['a0', 'a1', 'a2', 'a3', 'a4', 'a5', 'f'],
    combine="((; a0, a1, a2, a3, a4, a5, f), (x,)) -> a5()*(f(x))^5 + a4()*(f(x))^4 + a3()*(f(x))^3 + a2()*(f(x))^2 + a1()*(f(x))^1 + a0()",
)

model = PySRRegressor(
    expression_spec=template,
)
model.fit(X.reshape(-1, 1), Y)

Can you tell me how this can be fixed? Thank you very much!!!

[MilesCranmer]

Very weird, I’ll take a look.

By the way, you should avoid evaluating f(x) repeatedly if possible. Templates allow arbitrary Julia code so you can just make a variable that gets reused, like fx = f(x); a5()*fx + ... and so forth

[MilesCranmer]

Also, try out #823, should make such searches MUCH more efficient.

[MilesCranmer]

Moved to #829

[MilesCranmer]

Should be fixed on v1.4.0 of PySR once the backend v1.7.2 is available on the registry (~30 min).

[CleverCompiler-oss]

Everything is working! You've been very helpful! Thanks