Previous work on transfer learning focused on adapting solutions in a base domain to problems with similiar features or structure in a new domain. Different techniques are required for domains where problems are qualitatively dissimilar in both features and structures. In this work, we examine how transfer learning might be accomplished in the domain of "tower defense" spatial reasoning puzzles. Using a combination of human studies and generative computer models, we show that transfer is possible in this domain by using a set of strategies, possibly in a novel combination, inferred from multiple base problems.