This project we improve mini max search by prediciting which move the opponent is most likely to take. This prediction is performed by using a dynamic bayesian network. This network has nodes that represents a number of abstract features ex- tracted from the game. The network is trained using EM-learning based on data from a number of comptuter players playing against each other. These computer players are defined as the weighted sum of the features mentioned before. These weights are found using a genetic approach. The system is able to improve perfor- mance of the worst strategies with- out increasing the search space.