Learning Conventions via Social Reinforcement Learning in Complex and Open Settings

This article explores the computation of conventions in agents' societies via social learning methods, when agents aim to perform tasks collaboratively with their acquaintances. The settings considered require agents to accomplish multiple tasks simultaneously w.r.t. operational constraints, in coordination with their peers, even when they have limited monitoring and interaction abilities. In conjunction to that, agents have limited information about the potential strategies of others towards accomplishing tasks jointly, implying the need for coordination. The article formulates the problem as an MDP, and presents social reinforcement learning methods according to which agents interact concurrently with their acquaintances in their social contexts towards (a) learning about others' options and strategies to perform tasks, and (b) forming own strategies in coordination with others. The paper reports on the effectiveness of the learning methods w.r.t. to agents' constraints and limitations.