Evolutionary Feature Optimization

This repository contains a compact academic Python project focused on regression with evolutionary feature optimization. The code builds transformed feature spaces from tabular datasets, searches for useful feature subsets with evolutionary strategies or a genetic algorithm, and compares the optimized pipeline against baseline regressors.

The project documentation for the subject has been added to the repository.

Objective

The main objective is to evaluate whether an evolutionary search process can improve regression performance by selecting and synthesizing input features before fitting a linear model.

In the current implementation, the project:

• Loads a CSV dataset for a regression task.
• Cleans numeric data and optionally applies one-hot encoding to low-cardinality categorical variables.
• Generates candidate transformed features, including polynomial-style and interaction-based terms.
• Uses an evolutionary optimizer (ES or GA) to search for promising feature subsets.
• Trains a regularized linear model on the optimized feature set.
• Compares results against KNeighborsRegressor and RandomForestRegressor.
• Saves plots and result files for later inspection.

Repository Structure

.
├── data/
│   ├── AirfoilSelfNoise.csv
│   └── California.csv
├── vopt/
│   ├── __init__.py
│   ├── run_single.py
│   └── vopt_core.py
├── .gitignore
├── requirements.txt
└── README.md

Included Datasets

The repository currently includes two tabular datasets already used by the code:

• data/AirfoilSelfNoise.csv: regression dataset with target column SSPL.
• data/California.csv: regression dataset with target column MedHouseVal.

No additional dataset claims are made here beyond what is present in the repository.

Requirements

• Python 3.10+ is recommended.
• The project dependencies are listed in requirements.txt.

Installation

Create and activate a virtual environment, then install the dependencies:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

How to Run

From the repository root, run the experiment module with one of the bundled dataset configurations.

Default run:

python3 -m vopt.run_single

Run with the California dataset:

python3 -m vopt.run_single --dataset california

Run with the genetic algorithm instead of the evolutionary strategy:

python3 -m vopt.run_single --dataset airfoil --optimizer GA

Run with a custom CSV file:

python3 -m vopt.run_single --data-path /absolute/path/to/dataset.csv --target target_column_name

Optional flags:

• --output-dir /absolute/path/to/output_dir
• --use-one-hot
• --no-save

Expected Outputs

When output saving is enabled, the script writes results to outputs_single/ by default:

• fitness_curve.png
• dashboard_dispersion.png
• timestamped JSON summary
• timestamped optimized train/validation CSV files

The script also prints a summary to the console, including baseline MAE, optimized MAE, relative improvement, and a baseline leaderboard.

Technologies Used

• Python
• NumPy
• pandas
• scikit-learn
• Matplotlib

Notes on Reproducibility

• A fixed random seed (42) is used in the current code.
• The repository includes the datasets required for the two predefined runs.
• The code is designed for single-run experimentation rather than for packaging as a full reusable library.

Limitations

• The project is an academic experiment and still reflects that scope.
• Configuration is intentionally lightweight and only exposes a small command-line interface.
• The code currently focuses on regression datasets in CSV format.
• There is no automated test suite in the repository.
• Plots are shown interactively, which may behave differently depending on the local Python and Matplotlib environment.

Author

Yulicenia Diaz Cabrera

Academic project from the subject of biologically inspired artificial intelligence in Master's degree in Computer science and technology at Universidad Carlos III de Madrid.