On a new approach to the analysis of complex multi-armed bandits

A radically new approach to indexable systems pioneered by Bertsimas and Niño-Mora is utilised to provide novel analyses of classes of complex multi-armed bandits in which the individual bandits have their own decision structure. A new index result for an undiscounted model is established. Parallel server versions of the models are studied via (the dual of) an LP relaxation. This analysis yields a natural heuristic policy which is evaluated numerically. Copyright Springer-Verlag Berlin Heidelberg 1998