Network of Bandits

The distribution of machine learning tasks on the user's devices offers several advantages for application purposes: scalability, reduction of deployment costs and privacy. We propose a basic brick, Distributed Median Elimination, which can be used to distribute the best arm identification task in various schemes. In comparison to Median Elimination run on a single player, we showed a near optimal speed-up factor. This speed-up factor is reached with a near optimal communication cost. Experiments illustrate and complete the analysis: in comparison to Median Elimination with unconstrained communication cost, the distributed version shows practical improvements.