Teaching agents with human feedback: a demonstration of the TAMER framework

Abstract

Incorporating human interaction into agent learning yields two crucial benefits. First, human knowledge can greatly improve the speed and final result of learning compared to pure trial-and-error approaches like reinforcement learning. And second, human users are empowered to designate "correct" behavior. In this abstract, we present research on a system for learning from human interaction - the TAMER framework - then point to extensions to TAMER, and finally describe a demonstration of these systems.