Adaptive Epigenetic Neural Networks: A Biologically Inspired Proof-of-Concept for Neuroevolution

Abstract We propose a novel hybrid neuroevolutionary algorithm inspired by epigenetics that attempts to combine the best of natural selection and reinforcement learning. Traditional approaches like NEAT evolve topologies and weights. In this context, fixed topology refers to maintaining a constant network architecture across all individuals, with only neuron-level parameters (biases and epigenetic scalars) subject to adaptation. Our method, the Adaptive Epigenetic Neural Network (AENN), keeps weights fixed and evolves neuron-local epigenetic scalars that regulate activation strength. We demonstrate a proof-of-concept by achieving 97.90% test average accuracy on the Wisconsin Breast Cancer dataset with a small population of 30 individuals over 25 generations. This biologically inspired mechanism may offer a scalable, interpretable method for evolving fixed-topology networks.