Communication in Turn Based Multiplayer Games Using Deep Reinforcement Learning

Abstract

This work investigates communication in cooperative settings of multi-agent reinforcement learning. We look at what conditions make it easier or harder for meaningful communication to arise between the agents. This includes introducing and showing the usefulness of learning biases in a discrete or continuous setting. In order to do this we extend the game of Negotiation to a continuous setting, introduce a new environment called Sequence Guess, and introduce a new learning bias that helps facilitate the emergence of communication in a continuous setting.