Deep Tournament Selection (DTS) for Genetic Algorithms

Implementation of the paper "Deep Tournament Selection for Genetic Algorithms" (Eliad Shem-Tov, Ron Edri, Achiya Elyasaf — Ben-Gurion University of the Negev). 📄 Paper: link coming soon.

Installation

pip install eckity-dts

Public API

from eckity_dts import CachingEvaluator, DeepTournamentSelection, DTSPolicy

Constructing DTS

The policy combines a population encoder with a pointer network:

from eckity_dts import DeepTournamentSelection, DTSPolicy
from deep_tournament_selection.selection.population_to_vec_transformer import (
    PopulationToVecTransformer,
)
from deep_tournament_selection.selection.self_attention_pointer import (
    SelfAttentionPointer,
)

population_size = 100
vocab_size = 2  # maximum gene value + 1

encoder = PopulationToVecTransformer(
    vocab_size=vocab_size,
    emb_dim=32,
    latent_dim=32,
    n_heads=4,
    n_layers=2,
    dim_feedforward=256,
    max_pointers=population_size,
)

pointer = SelfAttentionPointer(
    pointer_len=population_size,
    d_model=32,
)

policy = DTSPolicy(
    pop_to_vec_transformer=encoder,
    pointer_transformer=pointer,
    device="cpu",
    train_every_n_gens=10,
    learning_rate=2e-3,
    final_lr=1e-3,
    epsilon_greedy=1.0,
    epsilon_greedy_decay=0.999,
    min_epsilon=0.2,
)

dts = DeepTournamentSelection(policy)

Use it as the EC-KitY selection method:

selection_methods=[dts]

Important parameters:

• population_size determines the rank-embedding and pointer-table capacities.
• vocab_size is the maximum integer gene value plus one.
• train_every_n_gens controls how often accumulated trajectories train the policy.
• epsilon_greedy and its decay control teacher-forced tournament selection versus learned selection.
• custom_reward_function, when supplied, receives current fitness, previous fitness, and population arrays.
• device can be "cpu" or "cuda" when a compatible PyTorch installation is available.

Fitness caching

EC-KitY may evaluate unchanged vectors again across generations. CachingEvaluator avoids recomputing fitness for vectors it has already seen:

from eckity_dts import CachingEvaluator

evaluator = CachingEvaluator(MyEvaluator())
print(evaluator.cache_stats())

Compatibility

• Python 3.10 or newer
• EC-KitY 0.4.x
• NumPy 2.0.2 or newer
• SciPy 1.13.0 or newer
• PyTorch 2.7.1 or newer
• overrides 7.7.0 or newer

These bounds are compatible with eckity-dnc, eckity-bert-ga, and eckity-bert-gp. None of the operator packages depends directly on another operator package.

Research repository

The repository contains the full paper experiments for Graph Coloring, Set Cover, and TSP, together with benchmark instances, notebook-style runners, and result figures. These research resources are not included in the eckity-dts wheel.

For repository development:

uv sync --extra dev --resolution lowest-direct
uv run pytest

Experiment entry points remain available from a source checkout:

python -m deep_tournament_selection.experiments.graph_coloring --instance queen8_12.col.txt --generations 200
python -m deep_tournament_selection.experiments.set_cover --instance scp41.txt --generations 200
python -m deep_tournament_selection.experiments.tsp --instance att48.tsp --generations 200
python run_experiments.py --selection both --runs 3 --generations 500

Results are written under runs/. Paper figures are stored under figures/, and the architecture diagram is under images/.

Development and release

uv run pytest
uv run ruff check .
uv build

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

License

This project is licensed under the BSD 3-Clause License. See LICENSE.