Proceedings of the Annual Meeting of the Cognitive Science Society

Explaining how intelligent systems come to embody knowl-edge of deductive concepts through inductive learning is afundamental challenge of both cognitive science and artificialintelligence. We address this challenge by exploring how adeep reinforcement learning agent, occupying a setting simi-lar to those encountered by early-stage mathematical conceptlearners, comes to represent ideas such as rotation and trans-lation. We first train a Dueling Deep Q-Network on a shapesorting task requiring implicit knowledge of geometric proper-ties, then we query this network with classification and prefer-ence selection tasks. We demonstrate that scalar reinforcementprovides sufficient signal to learn representations of shape cat-egories. After training, the model shows a preference for moresymmetric shapes, which it can sort more quickly than lesssymmetric shapes, supporting the view symmetry preferencesmay be acquired from goal-directed experience.