BERT Mutation for Genetic Algorithms

Overview

This repository implements the paper BERT Mutation: Deep Transformer Model for Masked Uniform Mutation in Genetic Algorithms.

The paper proposes a domain-independent mutation operator for genetic algorithms that uses a BERT-style masked model to predict beneficial gene replacements from context. To make this work for fixed-length GA representations, it adds an elite-guided data augmentation mechanism that creates additional learning signal from strong historical solutions.

In the paper, the method is evaluated on four domains: Frozen Lake, Artificial Ant, Graph Coloring, and Unweighted Set Cover. The reported results show faster convergence and better final fitness than standard mutation baselines and an adaptive operator-selection baseline, while maintaining meaningful population diversity.

Installation

pip install eckity-bert-ga

Using the operator

Import the public API from eckity_bert_ga:

from eckity.creators import GAIntVectorCreator
from eckity_bert_ga import BertMutation, EckityCustomMutation

Create the BERT mutation model and wrap it as an EC-KitY operator:

population_size = 100

creator = GAIntVectorCreator(
    length=INDIVIDUAL_LENGTH,
    bounds=(0, MAX_GENE_VALUE),
)

bert_mutation = BertMutation(
    max_int_val=MAX_GENE_VALUE,
    get_fitness_func=evaluate_vector,
    context_size=INDIVIDUAL_LENGTH + 1,
    mask_probability=0.1,
)

mutation_operator = EckityCustomMutation(
    mutation_operator=bert_mutation,
    population_size=population_size,
    probability=0.4,
)

Add mutation_operator to an EC-KitY subpopulation's operators_sequence. The BERT model is initialized locally from configuration and does not download pretrained model weights.

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
• Transformers 4.50.0 or newer

Results from the paper

Fitness by generation

The following figure presents representative best-individual fitness curves for BERT Mutation and the mutation baselines across the four evaluated domains.

Best-individual fitness by generation, averaged over 10 runs. Black dots mark the runtime cutoffs.

Statistical comparisons

The following table presents the paired exact permutation-test results comparing BERT Mutation against each baseline in every evaluated domain.

Holm-corrected p-values for comparisons between BERT Mutation and the baseline mutation operators.

Benchmark instances

Benchmark instance sizes used in the experiments. The individual length $L$ denotes the genome length optimized by the GA.

Domain	Instance	Problem size	Individual length $L$
Artificial Ant
Artificial Ant	aux_map1	$20 \times 20$, 91 food cells	283
Artificial Ant	aux_map2	$20 \times 20$, 69 food cells	286
Artificial Ant	john_muir	$32 \times 32$, 89 food cells	200
Artificial Ant	los_altos	$100 \times 100$, 157 food cells	800
Artificial Ant	santafe	$32 \times 32$, 89 food cells	400
Set Covering
Set Covering	scp41	200 rows, 1000 columns	1000
Set Covering	scp51	200 rows, 2000 columns	2000
Set Covering	scp52	200 rows, 2000 columns	2000
Set Covering	scp53	200 rows, 2000 columns	2000
Set Covering	scp54	200 rows, 2000 columns	2000
Set Covering	scp56	200 rows, 2000 columns	2000
Set Covering	scp57	200 rows, 2000 columns	2000
Set Covering	scp64	200 rows, 1000 columns	1000
Set Covering	scp65	200 rows, 1000 columns	1000
Frozen Lake
Frozen Lake	default	$8 \times 8$ grid, 64 states	64
Frozen Lake	rand10x10	$10 \times 10$ grid, 100 states	100
Frozen Lake	rand7x7	$7 \times 7$ grid, 49 states	49
Frozen Lake	rand8x8	$8 \times 8$ grid, 64 states	64
Frozen Lake	rand9x9	$9 \times 9$ grid, 81 states	81
Graph Coloring
Graph Coloring	games120	120 vertices, 1276 edges	120
Graph Coloring	myciel7	191 vertices, 2360 edges	191
Graph Coloring	le450_5a	450 vertices, 5714 edges	450
Graph Coloring	mulsol.i.2	188 vertices, 3885 edges	188
Graph Coloring	zeroin.i.1	211 vertices, 4100 edges	211
Graph Coloring	zeroin.i.2	211 vertices, 3541 edges	211

Getting started

Requirements

• Python 3.10 or newer
• A working PyTorch installation

The code depends on:

• numpy
• gymnasium
• torch
• transformers
• eckity

Installing for repository development

From the repository root:

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

Running the experiments

Artificial Ant:

python example_runner.py
python -m dnm_paper.experiments.artificial_ant

Frozen Lake:

python example_runner_frozen_lake.py
python -m dnm_paper.experiments.frozen_lake

Useful options:

python -m dnm_paper.experiments.artificial_ant --generations 100 --runs 3 --population-size 6
python -m dnm_paper.experiments.artificial_ant --maps-dir artifical_ant_maps --output-dir experiments/artificial_ant/runs
python -m dnm_paper.experiments.frozen_lake --generations 10 --runs 1 --population-size 100 --total-episodes 2000

By default, results are written under experiments/artificial_ant/runs/<map_name>/bert_mutation/<run_id>/results.json. Frozen Lake results are written under experiments/frozen_lake/runs/<instance_name>/bert_mutation/<run_id>/results.json.

The experiment modules, benchmark maps, datasets, and result figures are repository resources and are not included in the eckity-bert-ga wheel.

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

Project structure

dnm_paper/
  config.py                    Experiment configuration and default paths
  individuals.py               Custom ECKITY individual creator
  logging_utils.py             JSON statistics logger
  experiments/
    common.py                  Shared experiment helpers and mutation builder
    artificial_ant.py          CLI entry point and experiment orchestration
    frozen_lake.py             CLI entry point and experiment orchestration
  mutation/
    bert.py                    BERT-based mutation operator
    eckity_adapter.py          Adapter that plugs the mutation operator into ECKITY
  problems/
    artificial_ant.py          Artificial ant map loader and evaluator
    frozen_lake.py             Frozen Lake evaluator
    frozen_lake_instances.py   Named Frozen Lake benchmark instances
artifical_ant_maps/            Benchmark map files
pyproject.toml                 Package metadata and dependencies