Decision-Maker Alignment: Benchmark Datasets

Abstract

Translate Article

While alignment in artificial intelligence (AI) is broadly concerned with how AI systems align with human values, decision-maker alignment refers to how algorithms align with the values of individual decision makers. Consequently, the values targeted in decision-maker alignment are attributes that influence the decision-making process employed by humans. Example of a cognitive attribute is risk tolerance. The ideal environments for investigating decision-maker alignment are those in which the optimal decision may not be available, forcing humans to compromise and select a suboptimal decision. An optimal decision may not be available due to environment constraints such as limited resources, uncertainty, or some source of pressure. In contrast with supervised machine learning, decision-maker alignment can be investigated with labeled data in which decisions are made by different decision makers who are influenced by cognitive attributes. The problem is that datasets to study decision-maker alignment are difficult and expensive to create and they result in a small number of samples. For this reason, this paper proposes an approach to extend existing datasets for the purpose of studying decision-maker alignment. We exemplify our proposed approach by extending a dataset of health insurance alternatives.