Proceedings of the Annual Meeting of the Cognitive Science Society

Transfer learning is fundamental for intelligence; agents ex-pected to operate in novel and unfamiliar environments mustbe able to transfer previously learned knowledge to new do-mains or problems. However, knowledge transfer manifestsat different levels of representation. The underlying compu-tational mechanisms in support of different types of transferlearning remain unclear. In this paper, we approach the transferlearning challenge by decomposing the underlying computa-tional mechanisms involved in bottom-up associative learningand top-down causal schema induction. We adopt a Bayesianframework to model causal theory induction and use the in-ferred causal theory to transfer abstract knowledge betweensimilar environments. Specifically, we train a simulated agentto discover and transfer useful relational and abstract knowl-edge by interactively exploring the problem space and extract-ing relations from observed low-level attributes. A set of hier-archical causal schema is constructed to determine task struc-ture. Our agent combines causal theories and associative learn-ing to select a sequence of actions most likely to accomplishthe task. To evaluate the proposed framework, we compareperformances of the simulated agent with human performancein the OpenLock environment, a virtual “escape room” with acomplex hierarchy that requires agents to reason about causalstructures governing the system. While the simulated agent re-quires more attempts than human participants, the qualitativetrends of transfer in the learning situations are similar betweenhumans and our trained agent. These findings suggest humancausal learning in complex, unfamiliar situations may rely onthe synergy between bottom-up associative learning and top-down schema reasoning.