Inventing a Grounded Language by Playing Guess Who? – Creating a Communication Protocol with Deep Multi-Agent Reinforcement Learning

E. Jorge. Master thesis, Chalmers University of Technology and University of Gothenburg, supervisors M. Kågebäck and E. Gustavsson, examinator R. Jörnsten, November 2016.


Learning a new language from scratch is a hard task yet something that all children are tasked with. Grasping the concepts of communication without feedback or grounding concepts could pose a insurmountable task.

To understand how communication can emerge we propose to create an environment for a grounded protocol for communication to emerge. This is done by assigning multiple agents a cooperative task, playing the game \emph{Guess Who?}, such that images serve as grounding conversational topics in an interactive image search. The agents are equipped with \emph{Deep Recurrent Q-Networks} and our experiments show that the agents create an efficient protocol of communication for playing the Guess Who?. We also visualise the communication to show how the protocol is grounded in visual aspects of the images. Finally we show that varying the levels of noise in the communication channel is beneficial both in terms of learning speed but also in terms of capacity of the model.

Photo credits: Nic McPhee