Potential-based multiobjective reinforcement learning approaches to low-impact agents for AI safety

The concept of impact-minimisation has previously been proposed as an approach to addressing the safety concerns that can arise from utility-maximising agents. An impact-minimising agent takes into account potential impact its actions on state environment when selecting actions, so avoid unacceptable side-effects. This paper proposes and empirically evaluates implementation within framework multiobjective reinforcement learning. key contributions are a novel potential-based specifying measure impact, examination variety non-linear action-selection operators achieve acceptable trade-off between achieving agent’s primary task minimising environmental impact. These experiments also highlight unreported issue with noisy estimates for agents using action-selection, which broader implications application