Deep Reinforcement Learning of Abstract Reasoning from Demonstrations

摘要

Extracting a set of generalizable rules that govern the dynamics of complex, high-level interactions between humans based only on observations is a high-level cognitive ability. Mastery of this skill marks a significant milestone in the human developmental process. A key challenge in designing such an ability in autonomous robots is discovering the relationships among discriminatory features. Identifying features in natural scenes that are representative of a particular event or interaction (i.e. »discriminatory features») and then discovering the relationships (e.g., temporal/spatial/spatio-temporal/causal) among those features in the form of generalized rules are non-trivial problems. They often appear as a »chicken-and-egg» dilemma. This paper proposes an end-to-end learning framework to tackle these two problems in the context of learning generalized, high-level rules of human interactions from structured demonstrations. We employed our proposed deep reinforcement learning framework to learn a set of rules that govern a behavioral intervention session between two agents based on observations of several instances of the session. We also tested the accuracy of our framework with human subjects in diverse situations.