Aligning to Human Decision-Makers in Military Medical Triage
摘要
Expert human decision makers do not make optimal decisions in realistic domains; their decisions are affected by preferences, ethics, background experience, and contextual factors. Often there is no optimal decision, or any consensus on what makes a decision good. In this paper we consider the problem of aligning decisions to human decision makers. To this end, we introduce a novel formulation of an aligned decision-making problem, and present the Trustworthy Algorithmic Delegate (TAD), an integrated AI system that learns to align its decision-making process to target decision-makers using case-based reasoning, Monte Carlo simulation, Bayesian diagnosis, and Naturalistic decision-making. We apply TAD in a military triage domain, where experts make different decisions, and present experimental results showing that it outperforms baselines and ablations at alignment in this domain. Our primary claims are that the combined components of TAD allows for aligned decision-making using a small, learned case base and that TAD outperforms simpler strategies for alignment in this domain.