Value Alignment via Tractable Preference Distance
Value Alignment via Tractable Preference Distance
复制标题
通过易于处理的偏好距离进行价值调整
DOI：
10.1201/9781351251389-18
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
K. Venable
中科院分区：
文献类型：
--
作者：
Andrea Loreggia;Nicholas Mattei;Rossi;K. Venable
偏好在日常生活中无处不在：每当我们想做出选择我们最喜欢的替代品的决定时，我们都会使用自己的主观偏好。因此，计算机科学和人工智能中的偏好研究多年来一直非常活跃，具有重要的理论和实践成果[13，21]以及库和数据集[20]。在包括多智能体系统[25]和推荐系统[22]在内的许多场景中，用户偏好在驱动系统做出决策方面起着关键作用。因此，重要的是要有偏好建模框架，允许表达和紧凑的表示，有效的启发技术，高效的推理和聚合。如果我们希望人们信任人工智能系统，我们需要为这些系统提供区分人们通常所说的“好”和“坏”决策的能力。在许多情况下，决策的质量不应仅基于决策者的偏好或优化标准，还应基于与决策影响相关的属性，例如根据任何数量的外生来源给出的约束或优先级，决策是否符合伦理或法律的[5，24，26]。事实上，可能有具体的道德原则，这取决于上下文，可以而且应该凌驾于决策者的主观偏好。对于主观偏好，它们可能适用于复杂决策的一个或多个单独组成部分，而不是整个事情。例如，如果我们需要选择一辆车，我们可能会喜欢某些颜色而不是其他颜色，我们可能会喜欢某些品牌而不是其他品牌。我们也可能有条件偏好，比如如果汽车是敞篷车，我们更喜欢红色汽车。对于这些场景，CP-net形式主义[6]是一种方便且富有表现力的偏好建模方法，已在偏好处理社区[8，12，15，23]中广泛使用。CP网提供了一个有效的、紧凑的
Preferences are ubiquitous in everyday life: we use our own subjective preferences whenever we want to make a decision to choose our most preferred alternative. Hence, the study of preferences in computer science and AI has been very active for a number of years with important theoretical and practical results [13,21] as well as libraries and datasets [20]. In many scenarios including multiagent systems [25] and recommender systems [22], user preference play a key role in driving the decisions the system makes. Thus it is important to have preference modeling frameworks that allow for expressive and compact representations, effective elicitation techniques, and efficient reasoning and aggregation. If we want people to trust AI systems, we need to provide these systems with the ability to discriminate between what one would broadly call “good” and “bad” decisions. In many instances, the quality of a decision should not be based only on the preferences or optimization criteria of the decision makers, but also on properties related to the impact of the decision such as whether or not it is ethical or legal according to constraints or priorities given by any number of exogenous sources [5,24,26]. Indeed, there may be specific ethical principles, depending on the context, that could and should override the subjective preferences of the decision maker. For the subjective preferences, they may apply to one or more of the individual components of a complex decision, rather than to the whole thing. For example, if we need to choose a car, we may prefer certain colors over others, and we may prefer certain brands over others. We may also have conditional preferences, such as in preferring red cars if the car is a convertible. For these scenarios, the CP-net formalism [6] is a convenient and expressive way to model preferences that has been used widely in the preference handling community [8,12,15,23]. CP-nets provide an effective, compact CONTENTS