Deep Neural Crossover for EC-KitY

eckity-dnc provides the Deep Neural Crossover (DNC) genetic operator for EC-KitY.

DNC is described in “Deep Neural Crossover: A Multi-Parent Operator That Leverages Gene Correlations” (paper).

Installation

pip install eckity-dnc

Installing eckity-dnc also installs its EC-KitY, PyTorch, NumPy, and SciPy dependencies.

Usage

Import the public API from eckity_dnc:

from eckity.creators import GAIntVectorCreator
from eckity_dnc import (
    DeepNeuralCrossover,
    DeepNeuralCrossoverConfig,
)

Create the vector creator and DNC operator:

population_size = 100
individual_length = 160
number_of_gene_values = 161

individual_creator = GAIntVectorCreator(
    length=individual_length,
    bounds=(0, number_of_gene_values - 1),
)

dnc_config = DeepNeuralCrossoverConfig(
    embedding_dim=64,
    sequence_length=individual_length,
    num_embeddings=number_of_gene_values,
    running_mean_decay=0.95,
    batch_size=1024,
    learning_rate=1e-4,
    use_device="cpu",
    n_parents=2,
    epsilon_greedy=0.3,
)

dnc_operator = DeepNeuralCrossover(
    probability=0.8,
    population_size=population_size,
    dnc_config=dnc_config,
    individual_evaluator=your_eckity_evaluator,
    vector_creator=individual_creator,
)

Add dnc_operator to the EC-KitY subpopulation's operators_sequence. The supplied evaluator must inherit from EC-KitY's SimpleIndividualEvaluator and evaluate vectors created by individual_creator.

See dnc_runner_eckity.py for a complete bin-packing example using tournament selection and mutation.

Compatibility

• Python 3.10 or newer
• EC-KitY 0.4.x (tested with 0.4.1)
• PyTorch 2.7.1 or newer (tested with 2.7.1)
• NumPy 2.0.2 or newer (tested with 2.0.2)
• SciPy 1.13.0 or newer (tested with 1.13.0)

Development

Install the package and development tools, then run the tests:

python -m pip install -e ".[dev]"
python -m pytest

Release preparation and manual TestPyPI/PyPI commands are documented in RELEASING.md.

License

This project is licensed under the GNU General Public License v3.0. See LICENSE.