Efficacy of Deep Neuroevolution for Emergent Control of Multi-Agent Swarms

A functional evolutionary algorithm system for improving the behavior of pheromone-dropping “ant” agents in simulation is presented. Given a semi-degenerate set of initial conditions with biased food locations, the agents learn to go to locations where there will likely be food, but when food is distributed truly randomly the genetic algorithm is unable to find a solution. Traditional backpropagation-based RL methods are tested as alternate training options, but no neural network agent is able to approach the performance of a simple pre-programmed agent, suggesting that the foraging problem is fundamentally difficult to train a neural network decision agent for.