Evolution of communication through differentiation of communicative and goal-directed behaviors

Many studies have focused on the evolution of communication in artificial life, often presupposing dedicated communication channels such as voices or pheromones. However, it is plausible that only behavior existed when communication began and that some behaviors gradually evolved as a mean of communication. A study by Quinn has shown that communication can evolve purely through the behavior for two agents in a simulated environment. We extend this study to a more complex communication using three agents. In the simulation, the agents are equipped with two wheels and eight IR sensors and are controlled by identical neural networks. This network evolved by applying a genetic algorithm to a single population. After evolution, the agents’ behaviors had differentiated into communicative and goal-directed types. We analyze the differences between these two types of behavior quantitatively, and show that the evolved behaviors have communication properties such as mutuality, information flow, and invariance of the motion gain.