Encoding Strategies for Learning Geographical Concepts via Analogy

Creating systems that can work with people as apprentices, learning from them in natural ways, is a long-standing goal of AI. Being able to teach systems new concepts via sketching is an important step towards this goal. This paper explores the role of qualitative representations in learning geographic concepts indicated on 2D maps. We propose three principles for representation bias when learning via analogical generalization, and describe two dimensions of variation in qualitative encoding schemes for 2D maps. An experiment with multiple encoding schemes, using Freeciv, an open source strategy game, is described.