Adapting a Kidney Exchange Algorithm to Incorporate Human Values

Artificial morality is moral behavior exhibited by automated or artificially intelligent agents. A primary goal of the field of artificial morality is to design artificial agents that act in accordance with human values. One domain in which computer systems make such value decisions is that of kidney exchange. In a kidney exchange, incompatible patient-donor pairs exchange kidneys with other incompatible pairs instead of waiting for cadaver kidney transplants. This allows these patients to be removed from the waiting list and to receive live transplants, which typically have better outcomes. When the matching of these pairs is automated, algorithms must decide which patients to prioritize. In this paper, I develop a procedure to align these prioritization decisions with human values. Many previous attempts to impose human values on artificial agents have relied on the “top-down approach” of defining a coherent framework of rules for the agent to follow. Instead, I develop my value function by gathering survey participant responses to relevant moral dilemmas, using machine learning to approximate the value system that these responses are based on, and then encoding these into the algorithm. This “bottom-up approach” is thought to produce more accurate, robust, and generalizable moral systems. My method of gathering, analyzing, and incorporating public opinion can be easily generalized to other domains. Its success here therefore suggests that it holds promise for the future development of artificial morality.