The MASH project is a collaborative platform with the aim to experiment different methods in an unknown environment of large size. The application is a goal-planning task in a 3D video game where runs are expensive. Moreover, there is no prior knowledge, the decisions have unknown semantics, observations on the environment are partial and of big size and accomplishing the task by taking random decisions always requires a very long run. So, solving this task is a big challenge. In this paper, we extend Monte-Carlo Tree Search, which has been proved very effective for applications in which simulating is easy and fast, to contexts in which there are only "real" expensive runs. This generic approach combines Clustering and Monte-Carlo Tree Search.