Decentralized Multi-Armed Bandit Can Outperform Classic Upper Confidence Bound: A Homogeneous Case over Strongly Connected Graphs (2111.10933v3)

Abstract: This paper studies a homogeneous decentralized multi-armed bandit problem, in which a network of multiple agents faces the same set of arms, and each agent aims to minimize its own regret. A fully decentralized upper confidence bound (UCB) algorithm is proposed for a multi-agent network whose neighbor relations are described by a directed graph. It is shown that the decentralized algorithm guarantees each agent to achieve a lower logarithmic asymptotic regret compared to the classic UCB algorithm, provided the neighbor graph is strongly connected. The improved asymptotic regret upper bound is reciprocally related to the maximal size of a local neighborhood within the network. The roles of graph connectivity, maximum local degree, and network size are analytically elucidated in the expression of regret.