Partner-Aware Algorithms in Decentralized Cooperative Bandit Teams

When humans collaborate with each other, they often make decisions by observing others and considering the consequences that their actions may have on entire team, instead of greedily doing what is best for just themselves. We would like our AI agents to effectively in a similar way capturing model partners. In this work, we propose analyze decentralized Multi-Armed Bandit (MAB) problem coupled rewards as an abstraction more general multi-agent collaboration. demonstrate naive extensions single-agent optimal MAB algorithms fail when applied bandit teams. Instead, Partner-Aware strategy joint sequential decision-making extends well-known Upper Confidence Bound algorithm. analytically show proposed achieves logarithmic regret, provide extensive experiments involving human-AI human-robot collaboration validate theoretical findings. Our results partner-aware outperforms other known methods, human subject studies suggest prefer implementing strategy.