In constraint-based design, components are modeled by variables
describing their properties and subject to physical or mechanical
constraints. However, some other constraints are difficult to represent,
like comfort or user satisfaction. Partially defined constraints can be
used to model the incomplete knowledge of a concept or a relation. Instead
of only computing with the known part of the constraint, we propose to
complete its definition by using machine-learning techniques. Because
constraints are actively used during solving for pruning domains, building
a classifier for instances is not enough: we need a solver able to reduce
variable domains. Our technique is composed of two steps: first we learn a
classifier for the constraint's projections and then we transform the
classifier into a propagator. We show that our technique not only has good
learning performances but also yields a very efficient solver for the
learned constraint.