Multi-agent Reinforcement Learning for Swarm Retrieval with Evolving Neural Network

Abstract

This research investigates methods for evolving swarm communication in a simulated colony of ants using pheromone when foriaging for food. This research implemented neuroevolution and obtained the capability to learn pheromone communication autonomously. Building on previous literature on pheromone communication, this research applies evolution to adjust the topology and weights of an artificial neural network (ANN) which controls the ant behaviour. Comparison of performance is made between a hard-coded benchmark algorithm (BM1), a fixed topology ANN and neuroevolution of the ANN topology and weights. The resulting neuroevolution produced a neural network which was successfully evolved to achieve the task objective, to collect food and return it to a location.