Discovery of multivalued dependencies from database relations is viewed as a search in a hypothesis space defined according to the generalisation relationship among multivalued dependencies. Two algorithms for the discovery of multivalued dependencies from relations are presented. The top-down algorithm enumerates the hypotheses from the most general to more specific hypotheses which are checked on the input relation. The bottom-up algorithm first computes the invalid multivalued dependencies. Starting with the most general dependencies, the algorithm iteratively refines the set of dependencies to conform with each particular invalid dependency. The implementation of the algorithms is analysed and some empirical results are presented.