Performance of bandit methods in acoustic relay positioning

We consider the problem of maximizing underwater acoustic data transmission, by adaptively positioning a mobile relay. This is a classic exploration vs. exploitation scenario well-described by a multi-armed bandit formulation, which in its canonical form is optimally solved by the Gittins index rule. For an ocean vehicle traveling between distant waypoints, however, switching costs are significant, and the MAB with switching costs has no optimal index policy. To address this we have developed a strong adaptation of the Gittins index rule that employs limited-horizon enumeration. We describe autonomous shallow-water field experiments conducted in the Charles River (Boston, MA) with unmanned vehicles and acoustic modems, and compare the performance of different algorithms. Our switching-costs-aware MAB heuristic offers both superior real-time performance in decision-making and efficient learning of the unknown field.