Grounding abstraction in sensory experience

摘要

In order to make appropriate decisions, intelligent creatures narrow their sensory information down to high-level, abstract knowledge. Inspired by recent findings in neuroscience on the role of mirror neurons in action-based abstraction, in this study we propose a two-phase framework whereby a reinforcement learning (RL) agent attempts to understand its environment via meaningful temporally extended concepts in an unsupervised way. Throughout the process of concept extraction, the reinforcement learning agent makes use of a model of short-term and long-term memory to retrieve meaningful concepts from its environment. Empirical results by using e-puck robots reveal the capability of the proposed approach in retrieving meaningful concepts from environments. Moreover, simulation results show that the proposed approach can help an agent learn once and apply its knowledge in other environments which have similar structure without any further learning.