Collaborative Research: RI: Small: Modeling and Learning Ethical Principles for Embedding into Group Decision Support Systems

日常生活中的许多环境需要结合群体中个人的主观偏好来做出决定，比如去哪里吃饭，去哪里度假，雇谁，资助哪些想法，或者走哪条路线。在许多领域，这些主观偏好与适用于决策场景的道德价值、伦理原则或业务约束结合在一起，并且通常优先于偏好。无论是当人工智能系统向我们推荐产品时，还是当我们使用支持人工智能的系统做出群体决策时，都能敏锐地感受到道德价值观与主观偏好之间的潜在冲突。这项研究旨在通过提供机制来约束人工智能系统可以做出的决定，确保群体决策过程的结果与人类价值观保持一致，从而使人工智能更加负责。为了实现建立受道德约束的、支持人工智能的群体决策系统的目标，这个项目从人类那里获得灵感，人类经常根据来自道德、伦理或商业价值的一些外部优先事项来约束他们的决定和行动。这项研究项目将解决目前缺乏原则性的、正式的方法来将伦理嵌入到人工智能代理和支持人工智能的群体决策支持系统中，通过促进人工智能代理的安全性和健壮性的最新发展，鉴于人工智能广泛地触及我们的日常生活，这将对社会产生广泛的影响和好处。具体地说，该项目的长期目标是为群体决策系统在人工智能中嵌入伦理指导方针建立数学和机器学习基础。在机器伦理学领域，主要有两种方法：自下而上的方法侧重于数据驱动的机器学习技术，以及遵循符号和基于逻辑的形式主义的自上而下的方法。该项目通过三个具体目标将这两种方法更紧密地联系在一起。(1)伦理原则的建模和评估：本项目将使用知识表示和偏好推理方面的文献中的偏好模型来扩展社会选择理论和公平划分中的原则。(2)从数据中学习伦理原则：该项目将开发新的机器学习框架，以学习个人伦理原则，然后将其汇总用于群体决策系统。最后，(3)将伦理原则嵌入群体决策支持系统：该项目将为设计基于人工智能的伦理群体决策机制开发新的框架。这项研究将建立新的方法，使自上而下或基于规则的方法与自下而上或基于数据的方法在形式上和实验上统一，以便将伦理嵌入到群体决策系统中。该项目还将正式确立跨计算代理团队进行伦理和约束推理的框架。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Many settings in everyday life require making decisions by combining the subjective preferences of individuals in a group, such as where to go to eat, where to go on vacation, whom to hire, which ideas to fund, or what route to take. In many domains, these subjective preferences are combined with moral values, ethical principles, or business constraints that are applicable to the decision scenario and are often prioritized over the preferences. The potential conflict of moral values with subjective preferences are keenly felt both when AI systems recommend products to us and when we use AI enabled systems to make group decisions. This research seeks to make AI more accountable by providing mechanisms to bound the decisions that AI systems can make, ensuring that the outcomes of the group decision making process aligns with human values. To achieve the goal of building ethically-bounded, AI-enabled group decision making systems, this project takes inspiration from humans, who often constrain their decisions and actions according to a number of exogenous priorities coming from moral, ethical, or business values. This research project will address the current lack of principled, formal approaches for embedding ethics into AI agents and AI enabled group decision support systems by advancing the state of the art in the safety and robustness of AI agents which, given how broadly AI touches our daily lives, will have broad impact and benefit to society.Specifically, the long-term goal of this project is to establish mathematical and machine learning foundations for embedding ethical guidelines into AI for group decision-making systems. Within the machine ethics field there are two main approaches: the bottom-up approach focused on data-driven machine learning techniques and the top-down approach following symbolic and logic-based formalisms. This project brings these two methodologies closer together through three specific aims. (1) Modeling and Evaluating Ethical Principles: this project will extend principles in social choice theory and fair division using preference models from the literature on knowledge representation and preference reasoning. (2) Learning Ethical Principles From Data: this project will develop novel machine-learning frameworks to learn individual ethical principles and then aggregate them for use in group decision making systems. And finally, (3) Embedding Ethical Principles into Group Decision Support Systems: this project will develop novel frameworks for designing AI-based mechanisms for ethical group decision-making. This research will establish novel methods for the formal and experimental unification of aspects of the top-down or rule-based approach with the bottoms-up or data-based approach for embedding ethics into group decision making systems. The project will also formalize a framework for ethical and constrained reasoning across teams of computational agents.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.