For an autonomous system to provide value (e.g., to customers, designers, or society at large) it must have a reliable method to determine the intended goal. This is the essence of the value-alignment problem: ensuring that the objectives of an autonomous system match those of its human users. In robotics, value alignment is crucial to the design of collaborative robots that can integrate into human workflows, successfully learning and adapting to the objectives of their users as they go. We argue that a meaningful solution to the value-alignment problem will combine multi-agent decision theory with rich mathematical models of human cognition, enabling robots to tap into people's natural collaborative capabilities. We present a solution to the cooperative inverse reinforcement learning (CIRL) dynamic game using well-established models of decision making and theory of mind from cognitive science. The solution accounts for two crucial aspects of collaborative value alignment: that the human will not plan her actions in isolation, but will reason pedagogically about how the robot might learn from them; and that the robot should anticipate this and interpret the human's actions pragmatically. To our knowledge, this constitutes the first equilibrium analysis of value alignment grounded in an empirically validated cognitive model of the human.

智能系统与人的目标保持一致是价值对齐问题，我们提出一种基于多智能体决策理论和人类认知数学模型相结合的协作逆强化学习解决方案，该方案是价值对齐的首个基于经验证的认知模型的形式分析。

实用-教育价值取向对齐