NEAT Algorithm: Evolving Neural Networks Without Backprop

What Makes NEAT Special

Unlike standard NNs, NEAT starts with minimal networks and adds complexity only when needed. This prevents bloat and finds efficient solutions.

Core Concepts

Speciation

Networks are grouped into species based on structural similarity. Each species competes internally, preserving diversity.

Innovation Numbers

Every new structural mutation (new connection/node) gets a unique global innovation number. This allows meaningful crossover between different topologies.

Fitness Sharing

Fitness is divided by species size to prevent any one species from dominating.

Python Implementation (neat-python)

import neat

config = neat.Config(
    neat.DefaultGenome,
    neat.DefaultReproduction,
    neat.DefaultSpeciesSet,
    neat.DefaultStagnation,
    'config-feedforward'
)

def eval_genomes(genomes, config):
    for genome_id, genome in genomes:
        net = neat.nn.FeedForwardNetwork.create(genome, config)
        genome.fitness = run_game(net)  # your game/simulation

population = neat.Population(config)
population.run(eval_genomes, n=1000)