Comparative Evolutions of Swarm Communication

In this paper, we attempt to replicate an experiment by Marocco and Nolfi in which a grounded communication system was evolved in robot swarms. We hoped to develop bidirectional communication without any hard-coding by testing neural networks on a navigation task in which effective communication was essential to achieving the best score. In order to build a communication system, controllers would have to associate certain frequency values in an undifferentiated, continuous range with meaningful information. Building off of previous unsuccessful attempts to replicate this experiment, we evolved neural networks with fixed topologies in our own custom-built simulation. Overall, we achieved fitness scores comparable to the original experiment, although or recurrent networks performed significantly worse. A qualitative examination of our evolved populations, moreover, shows that we did not successfully evolve bidirectional communication. Frequency output was largely limited to binary values that aided with robot navigation. We conclude that our robots learned to effectively navigate the environment, but did not develop more complex communication behavior.