Cooperative Algorithms for Multi-Agent Multi-Armed Bandits: Integrating $\varepsilon$-Greedy Optimization

The multi-armed bandit framework is a wellestablished learning paradigm that enables sequential decisionmaking under uncertainty. This framework has been widely applied in various domains, including recommendation systems, clinical trials, online advertising, and operations research. In this paper, we consider a multi-agent multi-armed bandits scenario, where each agent possesses multiple arms and different agents may share common arms. In each round, an agent broadcasts its local observations to other agents that share the same arm. By combining two cooperative bandit algorithms—Cooperative Upper Confidence Bound Algorithm (CO-UCB) and Cooperative Active Arm Elimination Algorithm (CO-AAE)—we propose a novel cooperative algorithm that incorporates <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\varepsilon$</tex>-greedy optimization to balance exploration and exploitation. Through empirical evaluation, the proposed algorithm demonstrates effectiveness, owing to the similarity of CO-UCB total regret and CO-AAE computation time.