The capability of learning from experience is of critical importance in developing multi-agent systems supporting dynamic group decision making. In this paper, we introduce a hierarchical learning approach, aiming to support hierarchical group decision making where the decision makers at lower levels only have partial view of the whole picture. To further understand such a hierarchical learning concept, we implemented a learning component within the R-CAST agent architecture, with lower-level learners using the LogitBoost algorithm with decision stumps. The boosting-based learning agents were then used in our experiments to classify experience instances. The results indicate that hierarchical learning can largely improve decision accuracy when lower-level decision makers only have limited information accessibility.