[Re] No Press Diplomacy: Modeling Multi-Agent Gameplay

Daniel Ritter, Dylan Sam, Kevin Du, Shamay G Samuel, Cody West, Aaron Zhang

Abstract: Diplomacy is a strategic board game where different powers battle over control of supply centers in Europe. The original authors [1] developed supervised learning and reinforcement learning models to learn to play the No Press version of Diplomacy, beating the existing state of the art rule-based bots. The original paper utilizes various different machine and reinforcement learning techniques, including attention, encoder and decoder blocks, graph convolutional networks (GCN), LSTM, and FiLM [2]. Their implementation and code built off of extensive existing software frameworks like DAIDE [3], developed by the Diplomacy research community for interfacing with other bots. Furthermore, the authors have also developed a game engine that provides a simple interface for playing Diplomacy games. Because the authors of the paper released all their code for their models, the paper is not entirely comprehensive with their implementation details. Without being able to refer to their code, these ambiguities proved to make replication fairly difficult. We relied on communication with the paper authors in order to resolve a variety of ambiguities. Ultimately, this report details our attempts to reproduce the paper. We failed to reproduce the results for many reasons, including architecture ambiguities, expensive training times/compute resources required that were unmentioned in the original paper, and the complexity of this project given a 2-month time frame.

Track: Replicability

NeurIPS Paper Id: https://openreview.net/forum?id=B1ETuVrgUr&noteId=ByxN8Pfx_H