MultiNEAT

MultiNEAT is a portable library for neuroevolution - evolving neural networks with a genetic algorithm. It is based on NEAT (NeuroEvolution of Augmenting Topologies): networks start minimal and grow over the generations, and the population is divided into species so new structures get time to improve. The core is fast, dependency-free C++17 with Python bindings.

• NEAT - generational and steady-state (rtNEAT) evolution, speciation, phased searching
• HyperNEAT and ES-HyperNEAT - large geometric networks painted by an evolved CPPN
• Novelty search and behavior-based speciation
• Evolvable traits for neurons, links and genomes, including custom Python objects
• Recurrent networks, leaky integrator neurons, Hebbian plasticity, RTRL
• Everything is picklable - evaluate genomes in parallel processes
• A built-in 2D physics engine for simulated agents

Documentation (also in docs/)

Installation

pip install git+https://github.com/peter-ch/MultiNEAT.git

You need Python 3.8+ and a C++17 compiler; nothing else (no Boost). See Installation for details and for building the C++ library alone.

Example

import MultiNEAT as NEAT

params = NEAT.Parameters()
params.PopulationSize = 150
params.MinActivationA = 4.9          # steep sigmoids
params.MaxActivationA = 4.9
params.AllowLoops = False            # feed-forward networks
params.TournamentSelection = False   # truncation selection

# 3 inputs (x1, x2 and the bias), 0 hidden, 1 output
genome = NEAT.Genome(0, 3, 0, 1, False, NEAT.ActivationFunction.UNSIGNED_SIGMOID,
                     NEAT.ActivationFunction.UNSIGNED_SIGMOID, 0, params, 0)
pop = NEAT.Population(genome, params, True, 1.0, 0)

def evaluate(genome):
    net = NEAT.NeuralNetwork()
    genome.BuildPhenotype(net)
    error = 0
    for inputs, target in [([0, 0, 1], 0), ([0, 1, 1], 1), ([1, 0, 1], 1), ([1, 1, 1], 0)]:
        net.Flush()
        net.Input(inputs)
        for _ in range(3):
            net.Activate()
        error += abs(net.Output()[0] - target)
    return (4 - error) ** 2

for generation in range(300):
    genomes = NEAT.GetGenomeList(pop)
    for g in genomes:
        g.SetFitness(evaluate(g))
    best = max(g.GetFitness() for g in genomes)
    print(generation, best)
    if best > 15:
        break
    pop.Epoch()

The tutorial explains it step by step.

Documentation

Tutorial	evolving XOR, step by step
How NEAT works	genomes, speciation, reproduction, mutations, phased and steady-state search
Neural networks	activation, activation functions, Hebbian learning, RTRL
HyperNEAT and ES-HyperNEAT	substrates and CPPNs
Traits	evolvable properties of neurons, links and genomes
Novelty search	searching for novelty instead of fitness
Saving and parallel evaluation	files, pickling, multiprocessing
Visualization, Physics engine, C++
Parameters reference, API reference, FAQ

Examples

TestNEAT_xor.py	NEAT solving XOR
TestHyperNEAT_xor.py	HyperNEAT solving XOR
TestESHyperNEAT_xor.py, TestESHyperNEAT_xor_3d.py	ES-HyperNEAT on 2D and 3D substrates
NoveltySearch_simple.py	novelty search finding XOR without being asked to
NoveltySearch.py	novelty search vs. fitness search in a deceptive maze
ball_keeper.py	a robot learning to keep a ball in the air
TestTraits.py, TestCondTraits.py, PythonObjectTraits.ipynb	evolving traits
src/Main.cpp	XOR in pure C++

Contributing

See CONTRIBUTING.md. Changes are listed in the CHANGELOG.

License

GNU Lesser General Public License v3.0 - see LICENSE.

NEAT was developed by Kenneth O. Stanley and Risto Miikkulainen. MultiNEAT was written by Peter Chervenski with contributions from Shane Ryan and others.