Transparent Value Alignment

Abstract

As robots become increasingly prevalent in our communities, aligning the values motivating their behavior with human values is critical. However, it is often difficult or impossible for humans, both expert and non-expert, to enumerate values comprehensively, accurately, and in forms that are readily usable for robot planning. Misspecification can lead to undesired, inefficient, or even dangerous behavior. In the value alignment problem, humans and robots work together to optimize human objectives, which are often represented as reward functions and which the robot can infer by observing human actions. In existing alignment approaches, no explicit feedback about this inference process is provided to the human. In this paper, we introduce an exploratory framework to address this problem, which we call Transparent Value Alignment (TVA). TVA suggests that techniques from explainable AI (XAI) be explicitly applied to provide humans with information about the robot's beliefs throughout learning, enabling efficient and effective human feedback.