Neuroevolution

Neuroevolution is the application of evolutionary algorithms to the problem of designing neural networks — optimizing their weights, topology, or both through simulated natural selection rather than gradient descent. The approach was developed in the 1990s and enjoyed renewed interest with NEAT (NeuroEvolution of Augmenting Topologies, Stanley and Miikkulainen 2002), which evolves both network weights and architecture simultaneously by encoding topology in the chromosome and protecting structural innovations through speciation. Neuroevolution's principal advantage over gradient-based methods is that it does not require a differentiable objective function and can escape local optima through population diversity; its principal disadvantage is computational cost, since each candidate network must be fully evaluated during fitness scoring. Modern evolutionary strategies (OpenAI ES, 2017) have revived interest by demonstrating that gradient-free optimization can scale to large neural networks when parallelized across many workers, matching or exceeding reinforcement learning baselines on several benchmark tasks. The central limitation of neuroevolution as a model of biological neural development is the same as for genetic algorithms generally: fitness is externally specified, development is absent, and the evolutionary dynamics are far simpler than those of biological neural systems. Neuroevolution succeeds as engineering; its insights into how brains evolve are limited.