Design Principles for Creating Human-Shapable Agents.

摘要: In order for learning agents to be useful to non-technical users, it is important to be able to teach agents how to perform new tasks using simple communication methods. We begin this paper by describing a framework we recently developed called Training an Agent Manually via Evaluative Reinforcement (TAMER), which allows a human to train a learning agent by giving simple scalar reinforcement1 signals while observing the agent perform the task. We then discuss how this work fits into a general taxonomy of methods for human-teachable (HT) agents and argue that the entire field of HT agents could benefit from an increased focus on the human side of teaching interactions. We then propose a set of conjectures about aspects of human teaching behavior that we believe could be incorporated into future work on HT agents.