Transfer Learning in Programming by Demonstration

Programming by demonstration is an efficient way to acquire procedural knowledge, but it can require many examples to learn how to perform complex tasks. In this paper, we present a novel approach to using transfer learning in programming by demonstration. In particular, we describe a mechanism that learns how to transfer solutions to common subgoals from one task to another. We show that transfer learning leads to faster, more effective learning for programming by demonstration systems on later tasks in a sequence of related tasks. We compare performance of transfer-learning and traditional learning methods on two learning protocols: one in which the user explicitly identifies a shared task; and one in which the shared task is present, but not explicitly identified. Systematic experiments in a synthetic domain demonstrate the effectiveness as well as the scalability of our approach.