Ant-Neuroevolution

A Cornell project testing whether evolved neural networks can learn how ants forage. Real ants communicate indirectly by leaving pheromone for others to follow, a behaviour called stigmergy.

It simulates pheromone-dropping ant agents in PyTorch and improves them with an evolutionary algorithm, with standard reinforcement learning methods tried as alternatives. No neural network agent came close to a simple pre-programmed agent, which suggests foraging is hard to train a neural decision-maker for.