Does NEAT evolve weights as well as structure?

Yes; both are part of the original approach.

Why use innovation identifiers?

They help align corresponding genes during crossover between different network structures.

Does NEAT require differentiable fitness?

No. It can use fitness obtained from an external evaluation or simulation.

AI Glossary / AI and machine learning / Advanced

Neuroevolution of augmenting topologies

Definition

Neuroevolution of augmenting topologies, usually called NEAT, is an evolutionary method that searches for both neural-network connection weights and network structure. It begins with simple networks and can add nodes and connections as evolution proceeds.

What it means

Many neural-network workflows choose an architecture first and then train its weights. NEAT makes aspects of the architecture part of the search. Candidate networks compete according to a fitness measure, and variation creates new candidates.

The original method combines structural growth with mechanisms for aligning genes during crossover and protecting new structures while they improve. Historical markings identify corresponding connections, while speciation groups sufficiently similar networks for competition.

NEAT does not imitate biological evolution in every detail and is not a guarantee of discovering an optimal architecture. It is a particular search strategy whose effectiveness depends on the task, fitness function, population settings and evaluation budget.

How it works

A population of initially simple networks is encoded as genomes containing connection information and weights. Mutation can alter weights, add connections or insert nodes by modifying existing connections.

Historical innovation identifiers help crossover match corresponding genes between networks with different structures. A compatibility measure organizes networks into species, helping new structures avoid immediate elimination by already refined competitors.

Fitness determines reproduction and selection over successive generations. The algorithm needs many candidate evaluations, which may be costly or noisy. Constraints and validation remain necessary, especially when networks control actions in environments that differ from the training simulation.

Three practical examples

A game research project evolves a controller whose inputs describe the game state and whose outputs select actions.
A simulation study searches for compact control networks while allowing structure to grow when useful.
An optimization experiment compares topology-evolving networks with fixed-architecture neuroevolution under the same evaluation budget.

Think of it this way

Think of improving a small machine while occasionally adding a new component, rather than starting with a huge machine and adjusting every part. Newly modified designs receive some opportunity to be refined before competing directly with mature designs.

The analogy captures incremental growth and protection. It does not imply that every added part is useful or that increasing complexity necessarily improves performance.

Worked example

A structural mutation in a tiny network

Suppose a toy network has two input nodes and one output node. One connection runs directly from input A to the output with weight 0.8.

An add-node mutation disables that connection, inserts a hidden node H and creates connections A → H and H → output. In the original NEAT-style construction, the new incoming connection can start with weight 1 and the outgoing connection with the old weight 0.8.

This initialization aims to reduce disruption, but the exact behavior also depends on activation functions and other network details. It should not be described as guaranteeing identical outputs in every implementation.

The new structure receives innovation identifiers so later crossover can align its genes with related structures. Subsequent evolution can adjust the new weights. Whether the mutation helps is determined by fitness evaluations, not by the fact that the network has become more complex.

Illustrative case study

Fictional case: evolving a simulated controller

Redwood Games, an invented development team, tests NEAT for a simple simulated character. Fitness rewards reaching a target while penalizing collisions and excessive movement.

One evolved controller exploits a simulation quirk and receives a high score without behaving as intended. The team corrects the evaluation, tests across varied starting conditions and compares results with a simpler controller.

The case illustrates that evolutionary search optimizes the supplied fitness rather than the designer's unstated intention. It does not claim a real game-development gain or imply that a successful simulated controller is ready for physical deployment.

Common misconceptions and mistakes

NEAT is not simply backpropagation with an automatically chosen number of layers. Its core search is evolutionary, although later hybrid methods can combine ideas. Speciation is a computational diversity mechanism, not evidence that networks are living organisms.

A high fitness score does not prove general competence. The system can exploit weaknesses in the test environment or reward design. Structural growth also does not guarantee compactness in every run.

Business use

NEAT may be useful when a task has a measurable fitness signal and gradient information is unavailable, inconvenient or not the preferred approach. It is often explored in games, control simulations and evolutionary research.

Budget for repeated candidate evaluations and multiple random runs. Define constraints and failure penalties carefully, then test on conditions withheld from evolution. Compare with simpler methods under equivalent resources. Record genomes and evaluation settings so results can be reproduced. For real-world control, assess simulation transfer and operational safeguards separately from the evolutionary benchmark score.

Check your understanding

An evolved controller earns the highest reward by exploiting a simulator bug. Has NEAT failed to optimize its supplied objective?

Answer: Not necessarily. It may have optimized that objective very effectively. The evaluation failed to represent the intended behavior, so the fitness and validation process need revision.

Related terms

Genetic algorithms
Mutation
Fitness function
Neuroevolution
genetic algorithm
topology
crossover
speciation

Learn next

Learn next how fitness sharing and compatibility distance support diversity in NEAT. Then compare a fixed-topology evolutionary search with a topology-growing search using the same task, evaluation budget and held-out test conditions.

Sources

Primary source: Stanley and Miikkulainen, Evolving Neural Networks through Augmenting Topologies, https://nn.cs.utexas.edu/downloads/papers/stanley.ec02.pdf ; journal record, https://direct.mit.edu/evco/article/10/2/99/1123/Evolving-Neural-Networks-through-Augmenting . The original paper was consulted on connection genes, add-node mutation, historical markings and crossover; its stated purpose is to minimise initial disruption, not preserve identical outputs. Mutation example and case are original.

Previous: Neuro-fuzzy systems Next: Neuron