Add binary classification support via sigmoid_rmse fitness function

[Narden91]

Summary

This PR adapts SLIM_GSGP (and GP/GSGP) for binary classification following the approach of Bakurov et al. (2022).
It also fixes a NameError present in the original README example and restructures the feature into a proper first-class API.

---

Changes

New: binary_classification/ subpackage

• __init__.py
  Re-exports:

  • sigmoid_rmse
  • binary_sign_transform
    This allows convenient imports.

• example_binary_classification.py
  End-to-end working example using load_breast_cancer.

---

New fitness functions (evaluators/fitness_functions.py)

• sigmoid_rmse(scaling_factor)
  Factory returning a fitness function that:

  • Wraps tree outputs with a sigmoid
  • Computes RMSE
  • Bounds predictions to [0, 1] during training

• binary_sign_transform(y_pred)
  Converts raw outputs to binary labels at prediction time:

  • Negative → 0
  • Non-negative → 1

---

Config registration

Files updated:

• config/slim_config.py
• config/gp_config.py
• config/gsgp_config.py

Changes:

• sigmoid_rmse added to fitness_function_options in all algorithm configs.

---

API parameter

Files updated:

• main_slim.py
• main_gp.py
• main_gsgp.py

New parameter added to:

• slim()
• gp()
• gsgp()

sigmoid_scaling_factor: float = 1.0

Purpose:

• Allows tuning sigmoid steepness at call time
• Avoids modifying configuration files

---

README.md

Improvements and fixes:

• Replaced broken binary classification example:

  • Removed undefined binarized variable
  • Fixed incorrect sklearn import
  • Replaced wrong dataset

• Added clean working example

• Fixed incorrect log_path descriptions (previously all said "for slim")

• Removed duplicate copy_parent bullet

• Added sigmoid_scaling_factor to arguments table

• Updated Python version requirement to match setup.py (>=3.10)

---

setup.py

• Added missing dependency:

  • dill

---

Usage

from slim_gsgp.main_slim import slim
from slim_gsgp.evaluators.fitness_functions import binary_sign_transform

final_tree = slim(
    X_train=X_train,
    y_train=y_train,
    fitness_function="sigmoid_rmse",
    sigmoid_scaling_factor=1.0,
    slim_version="SLIM+ABS",
)

predictions = binary_sign_transform(final_tree.predict(X_test))

---

Reference

Bakurov, I., et al. (2022).
General purpose optimization library (GPOL).
Swarm and Evolutionary Computation, 68, 101028.

https://doi.org/10.1016/j.swevo.2021.101028