danuker/symreg

A Symbolic Regression engine

symreg's Issues

Pipenv

Use Python poetry: https://python-poetry.org/

Poetry seems to have 636/1661 (38.3%) of commits as bugfixes.

Deciding instead to use pipenv, which only has 1539/7001 (21.9%)

Should remove SolutionScore

It is a bit awkward to require SolutionScore just for passing it to ndim_pareto_ranking:

symreg/symreg/regressor.py

Line 56 in dfc3a3f

scores = SolutionScore.scores_from_dict(scores)

Use a list of dicts instead, and convert the methods to plain functions.

Postman

Use https://www.postman.com/ for easier collaboration
Make tests run automatically (some YAMLs i've heard)

Multiprocessing is slow

Multiprocessing is very slow, compared to evolving one generation.

One 50-individual generation evolves in around 10ms, whereas starting a new process is 100-200 ms.

Still, in benchmarks, small generations with frequent feedback are what work best.

There is no obvious way to parallelize, because the GIL would block the CPU-intensive threads.

Think about what can be done.

Add multiplicative mutation

In _new_leaf there is logic to add a Gaussian for optimizing constants.

Also add a multiplicative mutation choice (multiply the constant by a Gaussian).

This is useful while we don't have faster optimization routines implemented.

Clarify in readme that there is a pip package

Auto-detect arity

Use len(signature(add).parameters).

Update to Python 3.10

Currently, there is an error on Py3.10:

_______________________________________________________________________________________________________________ ERROR collecting tests/test_regressor.py ________________________________________________________________________________________________________________
ImportError while importing test module '/home/dan/projects/symreg/tests/test_regressor.py'.
Hint: make sure your test modules/packages have valid Python names.
Traceback:
/usr/lib/python3.10/importlib/__init__.py:126: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
tests/test_regressor.py:6: in <module>
    import symreg.ga
symreg/__init__.py:1: in <module>
    from .regressor import Regressor
symreg/regressor.py:4: in <module>
    from symreg.ga import GA
symreg/ga.py:6: in <module>
    from .nsgaii import nsgaii_cull
symreg/nsgaii.py:4: in <module>
    from orderedset import OrderedSet
../../.local/lib/python3.10/site-packages/orderedset/__init__.py:5: in <module>
    from ._orderedset import OrderedSet
E   ImportError: /home/dan/.local/lib/python3.10/site-packages/orderedset/_orderedset.cpython-310-x86_64-linux-gnu.so: undefined symbol: _PyGen_Send

Remove orderedset as a dependency, and use own implementation.

Support Scikit Learn Pipeline

Since SymReg works best with normalized data, it would be ideal to allow it to be used in a sklearn pipeline, with a StandardScaler(with_mean=False) preprocessor.

Make sure this is possible, and add an example to the README.

---

Alternatively (if Pipeline support needs sklearn dependency): have an argument to just normalize the data inside. Divide by the median of the data for the regression, then multiply back when returning the resulting functions.

Convert to Pytest fixtures

Start with make_seeded_regressor; learn it