Active preference learning to aid public decisions

当个人需要在选项之间进行选择时，比如是购买电动汽车还是混合动力汽车，他们必须首先确定这些选项的特征，例如价格或每加仑汽油的里程数，然后选择最符合他们偏好的选项。这可能是一项艰巨的任务，因为有许多选择，有复杂的方式来描述这些选择，当个人不确定如何在这些选择之间进行权衡时。决策辅助工具通过提供对可用选项的风险、成本和收益的简单描述，帮助个人制定这样的决策问题。然而，许多重要的决策涉及多个决策者，例如一个家庭购买一辆汽车，一群朋友选择看一部电影，甚至是公众成员为他们的城市、州或国家选择未来的能源政策。在这项研究中，我们将个人决策辅助推广为公共决策辅助，帮助不同类型的决策者群体利用有关个人和群体选择的信息达成共识。为此，我们将主动偏好学习的方法与将个人偏好映射到群体偏好的社会福利函数的使用相结合。我们的两个公共决策辅助工具1)通过向决策者询问最少的问题来准确地学习个人偏好来学习个人偏好，2)有效地学习群体社会福利函数，然后3)基于学习到的个人和群体偏好向群体做出推荐。通过这种方法，我们旨在回答三个研究问题：1)个人和群体在能源和环境政策中使用什么选择规则？2)什么主动学习方法可以最好地估计这些选择规则？3)社会偏好的异质性在多大程度上影响群体共识？这项研究通过结合行为决策研究、决策分析、主动机器学习和技术经济分析的理论和模型来推进决策的基本知识。该项目致力于研究公共决策援助方法的概念、方法和经验基础，以帮助利益相关者群体就公共政策达成共识。为此，该项目将主动偏好学习方法与社会福利优化相结合，前者选择信息量最大的选择集来学习偏好，后者基于群体行为学习从个人偏好到群体偏好的映射。三个目的推进了这项研究。AIM 1开发了一种新型的孪生神经网络体系结构，该体系结构可以主动学习决策者在许多不同类型的行为选择规则中的个人偏好，使用模拟和先前的数据来针对强大的基准测试该体系结构。目标2用齐次1惩罚扩展该体系结构，以从群体选择行为中学习群体社会福利函数，使用模拟来对照先导研究中建立的社会福利函数先验来测试神经网络。AIM 3在两种情况下收集新数据。第一个测试是在美国联邦能源政策的在线随机实验中测试最好的个人和群体主动偏好学习方法。第二个项目使用现场实验来帮助智利监管机构确定环境检查的优先顺序。这些结果扩大了对学习个人和群体偏好方法的能力和效率的科学理解，并帮助实践者使用最有效的方法在能源和环境公共政策方面达成群体共识。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。

When individuals need to choose between options, such as whether to buy an electric or hybrid car, they must first characterize those options, for example in terms of price or miles per gallon, then select the option that best satisfies their preferences. This can be a daunting task when there are many options, complex ways of characterizing those options, and when individuals are unsure about how to make tradeoffs among them. Decision aids help individuals formulate such decision problems by providing a simple characterization of the risks, costs, and benefits of the available options. Yet, many important decisions involve multiple decision-makers, such as a family purchasing a car, a group of friends choosing a movie to watch, or even members of the public choosing the future of energy policy for their city, state, or country. In this research we generalize the individual decision aid to a public decision aid, that helps groups of heterogeneous decision-makers come to consensus using information about individual and group choices. To do this, we combine methods from active preference learning with the use of social welfare functions that map individual to group preferences. Our two public decision aids 1) learn individual preferences by asking the minimum number of questions of a decision-maker to precisely learn preferences, 2) efficiently learn group social welfare functions, and then 3) make recommendations to groups based on the learned individual and group preferences. Using this approach, we aim to answer three research questions: 1) What choice rules do individuals and groups use for energy and environmental policy? 2) What active learning methods can best estimate those choice rules? 3) To what degree does heterogeneity in social preferences affect group consensus? The research forwards fundamental knowledge of decision-making by combining theories and models at the intersection of behavioral decision research, decision analysis, active machine learning, and techno-economic analysis. This project forwards research into the conceptual, methodological, and empirical foundations of a public decision aid approach for helping groups of stakeholders come to consensus on public policies. To do this, the project combines active preference learning methods that select the most informative choice sets to learn preferences, with social welfare optimization, that learns a mapping from individual to group preferences based on group behavior. Three aims advance this research. Aim 1 develops a novel twinned neural network architecture that can actively learn the individual preferences of decision-makers across many different types of behavioral choice rules, using simulations and prior data to test the architecture against strong benchmarks. Aim 2 extends that architecture with a homogeneous degree 1 penalty to learn group social welfare functions from group choice behavior, using simulations to test the neural network against social welfare function priors established in pilot research. Aim 3 collects new data in two contexts. The first tests the best individual and group active preference learning approaches in an online randomized experiment for US federal energy policy. The second uses a field experiment to help Chilean regulators prioritize environmental inspections. The results expand scientific understanding of the capability and efficiency of methods for learning individual and group preferences, and help practitioners use the most effective methods for reaching group consensus in energy and environmental public policy.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.