One important goal of machine learning is to create agents that can behave ethically. Such ethical behaviour would allow autonomous agents to be safely used in a wider range of cases, e.g. fully autonomous vehicles may encounter unexpected moral dilemmas during deployment. At the same time, there is no agreement within or between societies as to what constitutes ethical behaviour. Fortunately, there has been recent work in philosophy called ‘moral uncertainty’, e.g. here in Oxford by William MacAskill, that aims to address this very issue. In this talk, we first look at a paper by Ecoffet and Lehman (from Open AI) to see how ideas from moral uncertainty can be used alongside with Reinforcement Learning to implement autonomous agents with multiple ethical theories. We also look at my recent AAAI paper, which looks at fanaticism, a well-known problem within moral uncertainty. In particular, should small parts of society be able to completely determine what ethical behaviour means (such that the other parts of society are disregarded), just because they are morally repulsed? Consider that only a very small part of society believes that abortion is ethically unaccaptable. Due to fanaticism, agents may also consider abortions ethically wrong, despite the majority of society thinking otherwise; is this desirable? Finally, we look at interesting unsolved problems within moral uncertainty and their relevance to machine learning.
Bio
I am Jazon Szabo, currently a final year PhD student at the Safe & Trusted AI CDT, co-organised by King’s College London and Imperial College London. My main research interests are value alignment and machine ethics; I look at the use of moral uncertainty to address problems within these fields. I am also greatly interested in the field of AI Ethics: I have been running an online AI Ethics Reading Group for 3 years and I have helped create course material for an AI Ethics course at KCL.
