Graphical Bandits

Опубликовано 22 июня 2016, 18:37

We consider a setting for nonstochastic multiarmed bandits in which actions are vertices of a graph G, the edges of G denote similarities between actions, and the payoffs observed in each play are in the neighborhood of the played action. This setting interpolates between the standard bandit problem (where G is the empty graph) and the setting of prediction with expert advice (where G is the clique). I will describe simple extensions of the Exp3 algorithm applicable to undirected and directed graphs. The corresponding regret bounds are shown to scale with the independence number of G, yielding the known expert and bandit bounds as special cases.