“…By this we mean: constraints that can be imposed on features and structures without serious loss of expressive power; transformations of the featurediscovery problem to other tasks for which efficient algorithms are known; optimisation formulations that can be solved efficiently, learning and inferencing with structured output spaces and so on. We have pursued the following strategies: (a) pose the problem as a discrete optimisation problem and solve it heuristically [28,15,48,39,49], (b) pose the problem as a continuous (often convex) opti-misation problem with sparsity inducing regularizers and solves it optimally [27,36] and (c) study restrictions on the space of relational features and investigate empirically whether it is acceptable for a relational learner to examine a more restricted space of features than that actually necessary for the full statistical model [41,35,33] We have also looked at heuristics for speeding up inference algorithms in relational settings [34].…”