A Computational Analysis of the Necessity of Cue in Communication Emergence

Abstract

Understanding the communication emergent mechanisms is an important issue that is expected to have implications for AI technology. Computational models have attracted attention to compensate for the lack of archaeological evidence of communication emergence processes. However, many studies use environments in which communication necessarily emerges with small state spaces or no rewarding actions other than communication. Therefore, the emergence conditions in environments where agents can choose actions not directly related to communication have not been investigated enough. In this study, we focus on the concept of “cue,” which refers to the production of accidental signals that initiate communication emergence processes in the field of evolutionary biology, and design experiments to analyze the necessity of the cue by using a multi-agent reinforcement learning environment The experimental results show that the communication emergence by agents with the PPO algorithm is significantly affected by the existence of motivations that produce cue signals. In addition, it was also shown that the cue is necessary when the environment is sufficiently large, while it is not necessary when the environment is small.