An Efficient Dynamic Sampling Policy For Monte Carlo Tree Search