Recursive Monte Carlo Search for Bridge Card Play

Computer Bridge remains a challenging obstacle for Artificial Intelligence. For the last twenty years, the state-of-the-art playing programs have been using a depth-one Monte Carlo (MC) search approach, associated with an open card solver called Double Dummy Solver (DDS). When increasing the computing resources, the MC approach reaches a plateau, and its playing level cannot be improved. In this work, we study Recursive MC (RMC) for Bridge card play. We show that, with more computing resources, this approach performs better than MC. Rather than using DDS or any domain-dependent simulator, a level N + 1 RMC consists in using a level N RMC playing program as simulator, level-zero RMC being MC. This recursion mechanism can be iterated several times at the cost of increasing the computing time, each time a recursion level is added. This work focusses on card play with no trump in duplicate with either 13 cards per player or 5 cards per player. With 13 cards per player, level-one RMC is superior to MC with a margin of 0.5 trick per card distribution on average, which is statistically significant. This is the first time RMC is applied with success to computer Bridge card play.