Evolving Neural Networks through Augmenting Topologies

Summary

NEAT starts from a minimal I/O network and complexifies: innovation numbers align crossover across different topologies, explicit fitness sharing protects new structure, and incremental growth keeps the weight space small. XOR solved in 32 generations on average (2.35 hidden nodes). Double-pole with velocities 3,600 evals (vs ESP 3,800 / SANE 12,600). Double-pole no-velocity 33,184 evals vs ESP 169,466 and Cellular Encoding 840,000. Ablations show no-growth, no-speciation, random-init, and no-mating all hurt.

Keywords

neat · neuroevolution · speciation · topology · pole-balancing · genetic-algorithms · ut-austin

Topics

neuroevolution, topology search, genetic algorithms

Research notes

• Primary: UT Austin PDF of Evolutionary Computation 10(2):99-127, 2002 (DOI 10.1162/106365602320169811). Copyright 2002 Massachusetts Institute of Technology on the PDF; license field left blank (publisher copyright, not an open license stated on the posted file). Stanley/Miikkulainen, Department of Computer Sciences, UT Austin; kstanley/risto @cs.utexas.edu. No official code URL on the PDF. HF has no paper page (not an arXiv preprint). Discord posted nn.cs.utexas.edu PDF labeled NEAT paper. Benchmark is pole-balancing/XOR, not a new hosted corpus, so no datasets_local row.