Efficient and Robust Emergence of Conventions through Learning and Staying

In a multi-agent system (MAS), conventions serve as an effective mechanism to reduce frictions among agents and hence solve coordination problems. Convention emergence studies how agents’ behavior patterns give rise to conventions and how efficiently a convention forms. In a networked MAS, the question focuses on how conventions can arise when the agents’ positions are constrained. In this paper, we investigate convention emergence under the multi-player synchronous interaction model in networked MASs. In particular, we focus on the scenario that the agents is not informed the actions played by other agents, and the only information agents can perceive is whether an interaction is success or not. To facilitate the emergence of conventions, we propose a novel approach, namely Win-Stay-Lose-Learn (WSLL), to solve the problem of no observation and shorten the action transformation time when convention seeds conflict among agents. We conduct experiments to verify the robustness and effectiveness of our proposed method, experimental results show that our method outperforms other baseline approaches in terms of convergence speed under various circumstances.