From Genes to Neural Networks: Understanding and Building NEAT (Neuro-Evolution of Augmenting Topologies) from Scratch

Neuro-Evolution of Augmenting Topologies (NEAT) is an algorithm that dynamically evolves both the weights and the topology of neural networks, allowing their complexity to increase over generations. It addresses two major challenges in neuroevolution: how to perform crossover between structurally different networks and how to maintain diversity. NEAT solves these issues through two key innovations: historical markings, called innovation numbers, to align genes during crossover, and speciation to protect novel architectures. By dividing the population into species based on topological similarity, NEAT allows new structures to evolve without being immediately outcompeted by more established ones.