graphevol 0.0.2

Evolutionary graph-based optimization library with Python bindings

A scalable topology-evolution engine powered by a high-performance Rust backend.

Features

What is graphevol?

graphevol is a Python library for graph-based neuroevolution with Hierarchical Genetic Algorithm and memory-efficient evolutions, designed for fast and scalable evolutinary search.

The main class PyGenus plays roles in reproduction and selection in Generative Algorithm. PyGenus manages network structures and weights, and it runs repruduction, returning functions' strs and weights. It can adapt the strs major Python numerical libraries: Numpy, JAX and PyTorch. In selection, PyGenus recieves scores and updated weights of each set, and apply selective pressure stochastically based on each performance.

This project is currently ongoing, therefore the design can be dramatically changed by a major update.

---

Installation

pip install graphevol

Install Graphviz for visualization:

Linux

sudo apt install graphviz

macOS

brew install graphviz

---

Quick Start

import graphevol as ge

# Generate default YAML configs (please modify them to change GA behavior and operators)
ge.generate_default_settings("ga.yaml", "ops.yaml")

genus = ge.PyGenus(
    seed=0,
    ouput_dir="outputs",
    # environmental settings
    input_dim=4,
    output_dim=2,
    # network settings
    use_bias=True, # Use bias node
    sustain_cache=True, # Whether PyGenus uses a cache to track functions already used across generations.
    # configs
    ga_config_path="ga.yaml",
    ops_path="ops.yaml",
)

genomes: list[list[str, list[float]]] = genus.next_generation() # function's expression (str) and weights (list[float])
scores: list[list[float, list[float]]] = rollout_fn(genomes) # each score (float) and updated weights(list[float])
genus.selection(scores)