Learning Robustly Through Embedded Cognition

Abstract : In the ACT-R modeling approach we assume that in order to become skilled at a task, we start with knowledge that first consists of a set of memorized instructions that is interpreted by the cognitive system. This interpretation process is gradually internalized by a process we call production compilation, leading to improvements in performance in terms of speed and accuracy. The quality and robustness of the resulting skill partially depends on the representation of the initial instructions. We found that if these instructions are grounded in what is perceived in the world, the resulting performance is better in terms of speed and accuracy, but more importantly in terms of being able to deal with new and unexpected situations. The reason is that grounded instructions require less internal control, which reduces the computational complexity of the skill. Several studies with a simulated Boeing 777 Flight Management System have shown that participants who receive instructions grounded in perception perform much better than participants who receive classical checklist instructions.