Learning new concepts is critical to making sense of the world. Research on analogical reasoning suggests structure mapping and schema induction can enable discovery of new relational concepts. However, existing theories of schema induction and refinement are insufficient to explain acquisition of rich, compositional hierarchies of relational concepts. This paper offers a proposal for this sort of representation construction, founded on reinforcement learning to evaluate the predictive usefulness of higher-order relations, together with a mechanism of relational consolidation by which systems of relations (schemas) can be chunked into unitary entities. A computational model of these ideas is outlined and partially tested in simulations and human experiments. Implications and moderating factors for relational consolidation are considered.