Social Imitation in Cooperative Multiarmed Bandits: Partition-Based Algorithms with Strictly Local Information

We study distributed cooperative decision-making in a multi-agent stochastic multi-armed bandit (MAB) problem in which agents are connected through an undirected graph and observe the actions and rewards of their neighbors. We develop a novel policy based on partitions of the communication graph and propose a distributed method for selecting an arbitrary number of leaders and partitions. We analyze this new policy and evaluate its performance using Monte-Carlo simulations.