Collective Good and Optimization in Socioeconomic Systems

Daniel E. Rigobon’s May 2023 Princeton dissertation studies polarization in networks, liquidity choices among banks, and distributive objectives in supervised learning.^[1]

Method and contribution

It combines mathematical models, theoretical results, and simulations. Its fairness chapter defines a family of objectives between average predictive loss and worst-case loss, interpreted through utilitarian and Rawlsian perspectives. Dataset experiments examine tradeoffs and changes with model complexity.^[1]

Alignment relevance and limits

Social Alignment requires specifying whose outcomes count, rather than assuming an individual optimum is collectively beneficial. The fairness analysis also relates risk aversion to a modeled Veil of Ignorance. Mapping predictive error to ethical welfare is a substantive modeling assumption. Two distributive frameworks and selected datasets do not exhaust justice or establish a universally preferred objective. The dissertation identifies preprocessing, postprocessing, and other ethical theories as further work. Its network and banking conclusions depend on their respective model assumptions.