A BIOLOGICALLY INSPIRED METHOD FOR CONCEPTUAL IMITATION USING REINFORCEMENT LEARNING

Abstract

Learning by imitation can provide agents with a natural and effective means for transferring knowledge when brain-to-brain connection is infeasible. Natural mechanisms for imitation demonstrate strong abstraction and conceptualization capabilities, however, computational models that have been proposed for imitative learning barely address these fundamental features. Inspired by functions of human brain constituents and exploiting ideas enthused by mirror neurons and the multi-store model of memory, we propose a new model for learning by imitation capable of developing relational concepts. In our model, memory gradually organizes sensory data into concepts through reinforcement learning and consolidation, while mirror neurons maintain an extendible repertoire of familiar actions connected to corresponding concepts. We also discuss the relation between modeling behavior of concept-oriented agents in terms of mathematical functions and relevant biological evidence of mirror neurons. Eventually, we evaluate our method in a phoneme acquisition experiment through real interaction with humans.