neuroevolution

Neuroevolution is the application of evolutionary computation to the design and training of artificial neural networks. It evolves populations of networks to maximize fitness, potentially optimizing topology, connection weights, and sometimes learning rules. Neuroevolution is used in reinforcement learning, robotics, and pattern recognition, and is often presented as an alternative or complement to gradient-based methods such as backpropagation.

Networks are encoded as genomes that determine structure and parameters. A fitness evaluation runs the network on a task, and the outcome guides selection. Evolutionary operators include mutation and crossover, and strategies such as speciation help preserve diverse architectures. Some approaches promote modularity, while others search over larger, interconnected topologies.

One influential family is NEAT, NeuroEvolution of Augmenting Topologies. NEAT evolves both topology and weights from a minimal starting structure and uses speciation to preserve diverse solutions. It introduces mutations that add nodes and connections while keeping genomes compatible for crossover, aiding progressive complexity.

HyperNEAT extends NEAT with indirect encoding via compositional pattern-producing networks (CPPNs), enabling large-scale networks whose connectivity follows regular geometric patterns. Other methods use alternative encodings, such as Cartesian genetic programming, and may employ novelty search or multiobjective optimization to balance performance with simplicity and diversity.

Applications include game playing, autonomous robots, control systems, and optimization tasks. Neuroevolution can discover compact, generalizable architectures but is often computationally intensive and subject to bloat without proper regularization. Hybrid approaches combine evolutionary search with gradient-based fine-tuning to improve efficiency and performance.