Defeasible logic is a non-monotonic logic with applications in rule-based domains such as law. To ease the development and improve the accuracy of expert systems based on defeasible logic, it is desirable to automatically induce a theory of the logic from a training set of precedent data. Empirical evidence suggests that minimal theories that describe the training set tend to be more faithful representations of reality. We show via transformation from the hitting set problem that this global minimization problem is intractable, belonging to the class of NP optimisation problems. Given the inherent difficulty of finding the optimal solution, we instead use heuristics and demonstrate that a best-first, greedy, branch and bound algorithm can be used to find good theories in short time. This approach displays significant improvements in both accuracy and theory size as compared to recent work in the area that post-processed the output of an Apriori association rule-mining algorithm, w...