Scalar Reward Is Not Enough: A Response to Silver, Singh, Precup and Sutton (2021)

In a response paper, Peter Vamplew et al. rebut the scalar reward premise of Silver et al.'s "Reward is Enough," arguing for an explicit multi-objective reward model. The authors argue that scalar rewards cannot explain certain aspects of biological and computational intelligence; even if scalar rewards can trigger intelligent behavior in specific cases, using them for artificial general intelligence would still bring unacceptable risks of unsafe or unethical behavior.