What is NEAT (Neuroevolution of Augmenting Topologies)?

Question

I have looked up what NEAT is on youtube and the internet, but I can only find projects using NEAT, but apart from the wikipedia entry (which only says what it is in introduction, and is very confusing), I still have no idea what it is, is it a library, is it a type of neural network, is it a method of training neural networks? Sorry if this is an obvious question.

Answer 1

NEAT, or Neuro-Evolution of Augmenting Topologies, is a population-based evolutionary algorithm proposed by Kenneth O'Stanley [1].

The algorithm is based on various key characteristics:

Complexification

The networks in the original population are the easiest possible and the algorithm simply adds new architectural elements (neurons, connections). This means, the resulting networks tend to be very small.

Dodging opposing conventions by traditional markings

In conventional evolutionary algorithms it can simply happen that two individuals encode the identical behaviour but with very distinct genotype. This is called competing conventions. When such individuals are subjected to crossover, their children are expected to be worse than either parent. NEAT solves this by having traditional markings of new fundamental elements. When a new fundamental element is created, it is designated an innovation number. After that, when two individuals are crossed over, their genotypes are arranged in such a way that the identical innovation numbers match and only the different elements are replaced.

Speciation and fitness distribution

NEAT works with the theory of species. That is just a division of the community into several collections of individuals, called species. This division is based on the contrast of the individuals that is estimated based on the similar arrangement of their genotypes as is used when doing a crossover. The possibility of crossing over individuals from various species is then much shorter than crossover inside species. By raising the mating of more related parents, the children are less inclined to be much worse than the parents because the parents just were cooperative.

Also, inside the species, the fitness is assigned among the individuals. This serves two objectives:

(1) It shields individuals from changes - when a change happens, the fitness would usually be low but because there is fitness distribution, the individual has chance to optimize itself to adapt to this new fundamental change.

(2) Encourages heterogeneity because the larger the species, the more is the fitness shared and the less fit are the members of the species.