Concrete Problems in AI Safety

Abstract

The paper discusses potential accident risks from machine learning systems, categorizes them based on the source of the problem (wrong objective function, expensive objective function, or undesirable behavior during the learning process), and suggests research problems and approaches to address these issues.