Using a hierarchy of coordinators to overcome the frontier effect in social learning

We propose in this paper the use of a hierarchy of coordinators to improve the convergence of a network of agents to a global norm. A norm or a convention is an unwritten law that a society of agents agree on. Social norms are used by humans all the time. Choosing on which side of the road to drive a car and the right-of-way at an intersection are wellknown examples. In a multi-agent setting, a convention may refer to a dominant coordination strategy, a common communication language, or the right of way among a group of robots. Upon establishing a norm, the overhead of coordination drops and the reliability of the multi-agent system increases [2]. When studying the emergence of norms and conventions, we typically assume the interaction between agents is random: a pair of agents are selected randomly to interact with one another. The process repeats both concurrently (several pairs interact at the same time) and consecutively (each agent collects history of interactions). When agents are adaptive, the process is then referred to as social learning. The coordination game is perhaps the most widely used game for studying social learning as it presents an agent community with two equally plausible norms to choose from (i.e. two Nash equilibriums). It was shown that in the absence of any restriction on agent interactions, a norm is guaranteed to emerge in the simple social learning setting where agents play the coordination game [1].