Deep Reinforcement Learning of Abstract Reasoning from Demonstrations

摘要

We designed a Deep Q-Network (DQN) that learns to perform high-level reasoning in a Learning from Demonstration (LfD) domain involving the analysis of human responses. We test our system by having a NAO humanoid robot automatically deliver a behavioral intervention designed to teach social skills to individuals with Autism Spectrum Disorder (ASD). Our model extracts relevant features from the multi-modal input of tele-operated demonstrations in order to deliver the intervention correctly to novel participants.