Transfer as a Benchmark for Multi-Representational Architectures

We argue that transfer of spatial and conceptual knowledge between tasks and domains is an essential benchmark for multi-representational architectures aimed at human-level intelligence. The underlying hypothesis is that spatial relationships provide a natural level of abstraction, highlighting the similarities and differences between situations and domains. Therefore, not only will spatial representations improve domain reasoning and learning, they will also facilitate the transfer of knowledge across domains. The simulated environments of real-time strategy (RTS) games provide an excellent test-bed for exploring this hypothesis for two reasons: many different RTS domains have been constructed and RTS requires a wide range of reasoning tasks.