evozero

Evolve formulas and learning algorithms from zero — GPU-native evolutionary computation.

evozero is a GPU-accelerated evolutionary-computation toolkit with three engines that share one tensorized core:

• SymbolicRegressor — Eureqa-style symbolic regression: discovers an interpretable equation y = f(X). Island-model GP + tensorized postfix interpreter + linear scaling + constant optimization, returning a full complexity-vs-accuracy Pareto front.
• LearnerSearch — AutoML-Zero: evolves the learning algorithm itself (Setup/Predict/Learn over a typed-register VM), meta-trained across tasks so the discovered learner generalizes to unseen problems.
• launch_dashboard — a dependency-free (stdlib) live web dashboard of the search.

import evozero never imports torch: export utilities (LaTeX, NumPy) work on machines without a GPU. PyTorch is only imported when you actually run a search.

Install

pip install evozero                 # CPU (brings the CPU torch wheel)
pip install "evozero[cuda]" --index-url https://download.pytorch.org/whl/cu128   # NVIDIA GPU
uv sync --extra cuda                # from a clone, for development

evozero does not pin a CUDA build of PyTorch — you choose the wheel that matches your driver. See https://pytorch.org/get-started/.

Quickstart — symbolic regression

import numpy as np
from evozero import SymbolicRegressor

X = np.random.randn(500, 2)
y = X[:, 0] ** 2 + X[:, 0] * X[:, 1] + np.sin(X[:, 1])

model = SymbolicRegressor(population_size=3000, generations=200,
                          device="auto", random_state=0, verbose=1)
model.fit(X, y)

print(model.best_equation_)     # x0*(x0 + x1) + sin(x1)
print(model.to_latex())         # x_{0} \left(x_{0} + x_{1}\right) + \sin{x_{1}}
f = model.to_numpy_func()       # pure-NumPy callable, no torch needed
model.pareto_front_             # [{'complexity', 'loss', 'r2', 'equation'}, ...]

Quickstart — AutoML-Zero

from evozero import LearnerSearch, Task

tasks = [Task(Xi, yi) for Xi, yi in my_meta_training_data]
search = LearnerSearch(population_size=200, max_time=60, device="auto", random_state=0)
search.fit(tasks)

learner = search.export_learner()      # a discovered EvolvedLearner
learner.fit(X_train, y_train)
preds = learner.predict(X_test)
print(learner.to_python_source())      # the evolved learning algorithm, as code

Live dashboard

from evozero import launch_dashboard
with launch_dashboard(port=8080) as dash:
    print("open", dash.url)
    ...  # push metrics with dash.update({...}) from a training callback

or evozero dashboard --port 8080.

Documentation

https://evozero.readthedocs.io · Changelog · Contributing

Citing

If evozero helps your research, please cite it — see CITATION.cff.

License

Apache-2.0 © Karim Touma. See LICENSE.