A novel method is proposed to select materialized views of multi-dimensional data called dynamic selection. The idea of dynamic selection is that the system is in charged of collecting the queries to obtain their distribution. The set of materialized views is adjusted dynamically according to the queries?distribution. The method is given in detail including the algorithm of single-step selection and the instant adjusting method. It is also proved that under certain constraints, the performance of the single-step algorithm is guaranteed to be no worse than that of the optimal one by a certain bound.The xeperimental results show that the dynamic selection is more effective than other solutions.