geneticevo

Alpha preview: 0.1.0a1. APIs and checkpoint compatibility may change.

A Python library for reproducible evolutionary experiments, with optional lifetime learning and population tensor execution. Supply a problem and an agent factory; the runner handles selection, mutation, elites, lineage, evaluation, metrics, archives, and checkpoints.

Pure evolution requires only NumPy and Gymnasium. PyTorch, Stable-Baselines3 (SB3), PettingZoo, and plotting are optional. The package and Python import name are geneticevo. This project has not been published or deployed.

Evolution, learning, communication, and GPU execution do not guarantee better results or general intelligence. These are research tools, not deployment safety guarantees.

Installation

Python 3.10+ is declared; Python 3.11.9 was tested on Windows. From this directory:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e .
# Choose the extras you need:
python -m pip install -e ".[sb3]"       # PPO and DQN; includes PyTorch
python -m pip install -e ".[torch]"     # Native population tensors and PPO
python -m pip install -e ".[shared]"    # Official PettingZoo integration
python -m pip install -e ".[plot]"      # Saved plots
python -m pip install -e ".[dev,release]" # Test/build/metadata validation tools

On Linux/macOS, activate with source .venv/bin/activate. Source installation is the verified installation path; the project has not been uploaded to PyPI. To install the locally built alpha wheel, use:

python -m pip install "dist/geneticevo-0.1.0a1-py3-none-any.whl"

Core evolution does not install PyTorch, SB3, PettingZoo, or Matplotlib. Optional extras have their own upstream Python/platform requirements; for example, the tested PettingZoo 1.25.0 declares Python <3.13. Only Python 3.11.9 on Windows has been exercised here; the minimum declared Python version is not a tested matrix.

CPU and GPU requirements

• Ordinary CPU execution needs Python, NumPy, and Gymnasium. CUDA and a GPU are unnecessary for vector optimization and NumPy neural evolution.
• Native tensor inference/PPO requires the torch extra and works on CPU. SB3 learning requires the sb3 extra and also works on CPU.
• CUDA execution is experimental and hardware-unverified in this release. It requires an NVIDIA GPU, a compatible driver, and a compatible CUDA-enabled PyTorch build. Install the appropriate PyTorch build before requesting CUDA. The CUDA version string on an installed PyTorch build does not establish that a usable GPU exists.
• Python environments keep running on the CPU. GPU policies with Python environments are hybrid execution; native tensor simulation requires a compatible tensor environment. There is no general GPU speedup claim.
• An explicitly requested unavailable device raises an actionable error. Automatic CPU fallback is not used.

Declared ranges and tested versions:

Dependency	Declared range	Tested
NumPy	>=1.26,<3	2.4.4
Gymnasium	>=1.2,<1.3	1.2.3
PyTorch (optional)	>=2.12,<2.13	2.12.1+cu126, CPU execution
SB3 (optional)	>=2.9,<2.10	2.9.0
PettingZoo (optional)	>=1.24,<1.26	1.25.0
Matplotlib (optional)	>=3.10,<3.11	3.10.6

Only the listed versions were exercised; the entire allowed ranges were not tested. CUDA hardware was unavailable. See validation.

Complete quickstart

This environment is included and uses Gymnasium's reset/step conventions:

from geneticevo import Evolution, EvolutionConfig, GaussianMutation
from geneticevo.agents import MLPAgentFactory
from geneticevo.environments import Reach1D

experiment = Evolution(
    env_factory=Reach1D,
    agent_factory=MLPAgentFactory(hidden_sizes=(16, 16)),
    config=EvolutionConfig(
        population_size=12,
        generations=5,
        elite_count=2,
        selection="tournament",
        evaluation_episodes=3,
        seed=42,
        output_dir="results/quickstart",
    ),
    mutation=GaussianMutation(parameter_probability=0.1, std=0.1),
    learner=None,
    backend="serial",
    device="cpu",
)
result = experiment.run()
result.best_agent.save("results/quickstart/champion")
print(result.best_fitness)

The NumPy MLP has separate actor and critic networks. Its inherited vector contains all actor weights/biases, critic weights/biases, and, for continuous actions, log standard deviations. Critic genes matter when lifetime learning is enabled. Pure evolution does not use the critic for action selection.

Runnable examples

Run these from the repository after installation:

python examples/vector.py
python examples/neural.py
python examples/custom_genome.py
python examples/custom_learner.py
python examples/generalization.py
python examples/sb3_learning.py --algorithm PPO
python examples/sb3_learning.py --algorithm DQN --inheritance baldwinian
python examples/native_ppo.py --device cpu
python examples/native_ppo.py --continuous
python examples/shared_world.py
python examples/shared_world.py --rl

For CUDA, use python examples/native_ppo.py --device cuda. This path is implemented but hardware validation and performance are outstanding.

Configuration and lifecycle

generations=5 evaluates populations 0 through 4. There is no unused final offspring population. A generation initializes/restores, optionally measures pre-learning performance, learns, evaluates frozen policies, updates the archive, records statistics, and checkpoints. Reproduction happens only if another generation will be evaluated.

These settings have different meanings:

Setting	Meaning
population_size	Number of independently inherited individuals
generations	Number of evaluated populations
env_copies	Environment copies per individual, used by learners and tensor evaluation
learning_budget	Requested training transitions per individual per generation
evaluation_episodes	Frozen selection episodes per individual; also sizes the archive reference suite
Shared world's possible_agents	Roles acting together in one world; independent of population size

Serial evaluation visits episodes sequentially. CPU parallelism evaluates individuals in spawned processes. It requires pickleable top-level factories and an if __name__ == "__main__": entry point. Shared population evaluation is currently serial. SB3 learning runs individuals sequentially, even when fitness evaluation uses CPU workers.

Mutation probability applies per scalar parameter. std controls its magnitude. Tournament and rank selection honor maximize; custom selection, mutation, and optional crossover are supported. Crossover is disabled by default. immigrant_count injects fresh agents. diversity_pairs enables sampled parameter distances without an all-pairs calculation.

Target fitness and stagnation use fixed archive-reference fitness. Time and interaction caps stop at complete generation boundaries, so they can overshoot by a generation. The first population is always evaluated. Counts include archive and optional pre-learning evaluation. Custom vector objectives consume zero environment interactions.

Learning and inheritance

Use learner=None and learning_budget=0 for evolution only.

• Lamarckian: learned parameters replace inherited genes before reproduction.
• Baldwinian: train a fresh copy; fitness and the deployed archive snapshot use that copy, while reproduction uses the original genes.
• Unchanged elites continue optimizer/counter/replay state by default and learn again (elite_learning=True). Set it false to skip further elite learning with serial SB3/custom learners. Native batching requires it true.
• Mutated children and immigrants start fresh optimizer, replay/rollout, exploration, counter, and random-action state. DQN targets are synchronized to the child's mutated online network. Unchanged elites preserve DQN targets.
• Environments reset at generation boundaries. Mid-episode state is not resumed.

SB3's factory builds the algorithm's own compatible MlpPolicy; arbitrary NumPy or PyTorch modules cannot attach to it unchanged. PPO evolves its full online policy, including value and distribution parameters. DQN evolves its online Q network; target networks are algorithm state. Normalization wrappers are rejected until a frozen-statistics serialization adapter exists.

PPO rounds a budget up to rollout_steps * env_copies (SB3 uses n_steps). Updates may occur within unfinished episodes. DQN's default step frequency is one; custom frequencies may overshoot. Requested and actual counts are distinct. With SB3, set the factory's and learner's env_copies to match the config.

Native PPO batches independent policies, losses, gradients, clipping, and Adam. Adam has separate moments and step counters for each individual, so a fresh child can learn alongside an older elite. Advantages normalize per individual over a rollout. Discrete categorical and continuous tanh-squashed Gaussian likelihoods are implemented. Continuous entropy regularization uses the base normal's entropy; its default coefficient is zero.

Supported combinations

Agents / learner	Serial CPU	CPU process evaluation	Batched policies, Python env	Tensor environment
Vector objective / none	Yes	Yes	No	No
NumPy MLP / none	Yes	Yes	Yes	Yes
SB3 PPO or DQN / SB3Learner	Yes	Yes; learning stays sequential	No	No
MLP / NativePPO	No	No	Yes	Yes
Evolved message head / none	Shared serial	No	No	No
SB3 task policy / fixed messages	Shared serial	No	No	No

Neural observations must be flat Box; actions must be zero-start Discrete or flat, finite, strictly bounded Box. DQN supports Discrete only. Shared worlds use the PettingZoo Parallel protocol with flat Box observations. Other spaces need explicit tested adapters.

backend="batched" uses fixed shared architectures. Tensor computation also runs on CPU. Omitting tensor_env_factory uses GymTensorBridge: Python simulation stays on CPU even if policies run on CUDA. This is labeled hybrid execution. Setting a device never migrates arbitrary Python simulation code. Native tensor environments require matching policy/environment devices. The included TensorReach1D has a matching CPU reference. Archives always use the serial CPU reference evaluator, including for tensor runs.

Shared worlds and communication

Communication is disabled by default. MessageChannel uses bounded fixed-size messages and broadcasts to other active agents. Each sender owns an identified inbox slot plus a presence bit. Messages sent with a task action become visible in the next observation. Reset clears inboxes. Optional L1 message cost is subtracted from sender reward. Recipient and transform hooks allow custom routing/noise/dropout; queued delays need a custom channel implementation.

Policies receive only their permitted local observation and prior inbox. Pure evolution can change both task and message heads. SharedEvaluator rotates focal roles and current-population partners for selection; archive comparisons use a fixed partner pool. Use another evaluator with unfamiliar partners for final testing. reward_mode chooses focal-individual or mean-team reward. World steps and all active-agent transitions are counted separately; inactive roles do not act. Explicit focal-to-role mappings are retained in evaluation details.

The tested multi-agent learning integration is PPO/DQN on SharedFocalEnv against frozen partners with fixed message rules. The learner trains task actions; messages have no trainable distribution. End-to-end learned stochastic messages and simultaneous RL of distinct communicating policies are unsupported. Experience sharing, parameter averaging, imitation, and global memory are not part of communication.

Results, checkpoints, and generalization

output_dir writes metadata.json, incremental metrics.jsonl, lineage.json, and atomic latest.checkpoint. Metrics include fitness quantiles, task metrics, episode lengths, training/evaluation costs, learner reports, population and mutation settings, component timings, throughput, and cumulative wall time. Final-info success/failure fields are retained when environments provide them. Trajectories and every network are not saved by default.

The archive shortlists the best archive_size current candidates, then compares them with old archive snapshots on common, fixed reference scenarios. It does not compare incomparable rotating selection scores across generations. Snapshots are independent copies with their evaluation context.

Resume at an evaluated generation boundary:

# Re-create experiment with the same components/configuration; generations may increase.
result = experiment.run(resume_from="results/quickstart/latest.checkpoint")

Checkpoint schema 1 stores population/archive genes, resumable learning state, RNG state, IDs, lineage, and counters. Scalar settings and component descriptions are checked on restore. Custom components should expose configuration as public fields on stable classes; arbitrary hidden global or closure state cannot be verified. Load only trusted checkpoints: they use Python pickle. Agent folders save inference genomes; continuing optimizer/replay state requires a checkpoint. Bitwise reproducibility across devices, versions, and platforms is not promised.

Training uses a separate training_env_factory (defaults to env_factory). Selection scenarios rotate by generation, with a common suite for each population. Archive scenarios stay fixed. validation_sampler and archive_sampler accept (rng, count, generation) and return Scenario objects. Scenario.options reaches Gymnasium reset, enabling procedural distributions.

Final testing is explicit and never drives the runner:

from geneticevo.core import scenarios_for
from geneticevo.evaluation import GymEvaluator
test = result.evaluate(GymEvaluator(Reach1D), scenarios_for(2026, "test", 20))

Use separate test scenarios and partners, including shifted distributions. adaptation_evaluation gives every task a fresh training-state copy and records pre/post performance and costs. It does not modify the champion. Repeated training seeds and uncertainty_summary describe uncertainty across runs; additional episodes do not replace independent training seeds. No mechanism prevents memorization or invalid experimental design.

Development and evidence

python -m unittest discover -s tests -v
python -m pytest -q
python -m build --wheel --no-isolation
python -m twine check --strict dist/geneticevo-0.1.0a1*
python benchmarks/equal_work.py --repeats 3
python benchmarks/baselines.py --seeds 1 2 3

Windows sandbox restrictions may prevent spawned worker pipes; use an ordinary local terminal for process tests. Architecture and extensions explain custom genomes/environments/learners. Validation documents correctness checks and benchmark methodology. Delivery report records measured results and remaining limitations. The alpha release checklist records packaging checks and the boundary between CPU validation and unverified CUDA execution.

For a saved plot: from geneticevo.plotting import plot_history, then call plot_history("results/neural/metrics.jsonl", "results/neural/fitness.png").

License

geneticevo is licensed under the MIT License. Copyright notices and the license text ship with the source distribution and wheel. Dependencies retain their own licenses.