Operators¶
Predefined¶
First, note that pretty much any valid Julia function which takes one or two scalars as input, and returns on scalar as output, is likely to be a valid operator^{1}. A selection of these and other valid operators are stated below.
Binary
+

*
/
^
max
min
mod
cond
 Equal to
(x, y) > x > 0 ? y : 0
 Equal to
greater
 Equal to
(x, y) > x > y ? 1 : 0
 Equal to
logical_or
 Equal to
(x, y) > (x > 0  y > 0) ? 1 : 0
 Equal to
logical_and
 Equal to
(x, y) > (x > 0 && y > 0) ? 1 : 0
 Equal to
Unary
neg
square
cube
exp
abs
log
log10
log2
log1p
sqrt
sin
cos
tan
sinh
cosh
tanh
atan
asinh
acosh
atanh_clip
 Equal to
atanh(mod(x + 1, 2)  1)
 Equal to
erf
erfc
gamma
relu
round
floor
ceil
sign
Custom¶
Instead of passing a predefined operator as a string, you can just define a custom function as Julia code. For example:
PySRRegressor(
...,
unary_operators=["myfunction(x) = x^2"],
binary_operators=["myotherfunction(x, y) = x^2*y"],
extra_sympy_mappings={
"myfunction": lambda x: x**2,
"myotherfunction": lambda x, y: x**2 * y,
},
)
Make sure that it works with
Float32
as a datatype (for default precision, or Float64
if you set precision=64
). That means you need to write 1.5f3
instead of 1.5e3
, if you write any constant numbers, or simply convert a result to Float64(...)
.
PySR expects that operators not throw an error for any input value over the entire real line from 3.4e38
to +3.4e38
.
Thus, for invalid inputs, such as negative numbers to a sqrt
function, you may simply return a NaN
of the same type as the input. For example,
would be a valid operator. The genetic algorithm will preferentially selection expressions which avoid any invalid values over the training dataset.

However, you will need to define a sympy equivalent in
extra_sympy_mapping
if you want to use a function not in the above list. ↩