Title: Shrewd Selection Speeds Surfing: Use Smart EXP3!

Abstract: Mobile devices often have several wireless networks at their disposal. Choosing the right one is vital for good performance. Yet, solving the network selection problem in a distributed and fair manner remains a challenge. The contribution of this work is twofold. First, we formulate the problem as a repeated congestion game and model the behavior of the devices using online learning in an adversarial bandit setting. We study the performance of EXP3, a standard bandit algorithm, in a dynamic wireless network setting. We observe that it is relatively poor. Second, we propose Smart EXP3, an algorithm with good theoretical and practical performance. We prove that it has the same convergence and regret properties as EXP3 and evaluate it for the wireless network selection problem. Empirical results show that it gracefully deals with transient behaviors, has relatively fast convergence to the optimal network with reduced switching, and achieves fairness among devices. In real world experiments, Smart EXP3 can achieve 1.2x faster download over alternate strategies. More broadly, this paper provides a roadmap for how to adapt and use adversarial bandit algorithms (which are beautiful in theory, and yet often do not work well in practice) to efficiently solve networking problems.