Case-based reasoning (CBR) is a problem solving technique that puts at work the general principle that similar problems have similar solutions. In particular, it has been proved effective for classification problems. Fuzzy set-based approaches to CBR relyon the existence of a fuzzy similitary functions on the problem description and problem solution domains. In this paper, we study the problem of learning a global similarity measure in the problem description domain as a weighted average of the atribute-based similarities and, therefore, the learning problem consists in finding the weighting vector that minimizes mis-classification. The approach is validated by comparing results with an application of case-based reasoning in a medical domain that uses a diferent model.