Automated Negotiation in Supply Chains A Generalist Environment for RL/MARL Research

Yasser Mohammad, Shinji Nakadai, Amy Greenwald

Abstract: Reinforcement learning has been used to create competitive strategies for a variety of games ranging from zero-sum two-player complete-information games to multiagent incomplete information games. Most of these advances were achieved in computer and board games with clear-cut rules and environmental contexts. Extending these successes to business applications proved to be tricky in many cases because of the complexities introduced by the real world including scalability problems, changing game rules, and preference elicitation issues. In this paper, we propose a new game that can be used to benchmark RL and MARL methods. This game involves automated negotiation in a supply-chain management context.