Logographic Emergent Communication In A Multi-Agent Sketching Lewis Game

A recent development in the field of emergent communication has been communication by sketching. However, the communication protocols that emerge are inclined to create abstract sketches of images, rather than creating a symbolic communication protocol based on the concepts themselves. In this work, agents play a variation of the Lewis game with agents communicating using a limited sketching toolset, in an attempt to establish an emergent logographic character set. We also propose the notion of logographic convergence within a visual communication system, which we use to evaluate whether emergent communication protocols converge into a logographic character set. We compare two learning methodologies, reinforcement learning and end-to-end differentiable, for generating sketches, and compare their communication performance alongside evaluating whether logographic convergence is reached through the metrics of symbolicity and semanticity.