Miles Cranmer

[Feature] Reusing parts of equation

11

PySR should be able to reuse parts of an equation. In theory, a PySR equation (`Node`) type could have a single subtree used by multiple parts of the expression, as...

[Feature] Faster allocations during tree evaluation

1

Can probably speed up the tree evaluation with these clever suggestions from @mkitti: https://discourse.julialang.org/t/faster-zeros-with-calloc/69860

[Feature] Use NaNMath instead of `abs`

4

Operators should be default use NaNMath https://github.com/mlubin/NaNMath.jl, instead of my versions which simply map input to the valid domain like `log(abs(x))`. Since the evaluator already detects NaNs by default, this...

[Feature] TensorSymbolicRegression.jl

4

For example, say you would like to search for partial differential equations with genetic algorithms. You would define operators such as finite difference gradients, laplacians, or summations–and then search for...

[Feature] Callable trees

1

It might be more intuitive to call an expression like so: ```julia options = Options(binary_operators=(+, -, *)) x1, x2 = Node("x1"), Node("x2") tree = x1 * x2 - x1 X...

[WIP] Optimize expressions with Zygote-based autodiff

2

cc @kazewong This is an attempt to use `evalGradTreeArray` in the constant optimization. This function computes the explicit Zygote.jl-based derivatives of an expression with respect to its constants, so _in...

Right now, only operators which are already in `Base` are defined for `::Node` type automatically when declaring `Options`: https://github.com/MilesCranmer/SymbolicRegression.jl/blob/966026391b809f4642e7a93d873ee986737293ed/src/Options.jl#L446-L453 I'm not sure if this is necessary. It would be great...

[Feature] Increase Abstractions

1

I would like for SymbolicRegression.jl to have a mode that is much more abstract in terms of what operators are allowed. e.g., vector/matrix, or even symbolic operations. This would enable...

[Feature] Differential Operators

1

It would be great if SymbolicRegression.jl supported differential operators. To implement this with minimal changes, you could have a single operator for each variable. For example, `unary_operators=(dx[1], dx[2])` would create...

In the current code, using `batching=true` doesn't seem to get nearly as good a speedup as simply inputting a smaller dataset. This indicates either the random slicing is taking too...