Icarus, Towards Diplomatic Agents in Diplomacy

Diplomacy is a 7-player game where players (autonomous agents or human players) compete for territories but need teamwork and eventual betrayal to get an edge over their opponents. Agents in Diplomacy often neglect diplomatic interaction with other players, which is a missed opportunity in terms of creating interesting agents to play against. This paper explores how to create and train an agent that is able to perform diplomatic interactions with other players, such as proposing alliances. Towards this end, we explore recent deep learning techniques and architectures such as Transformers and LSTMs, and propose the Icarus architecture, a deep neural network with an explicit Diplomatic State representation.