Preference-based Constrained Reinforcement Learning for Safety

The recently introduced Preference-based Constrained Reinforcement Learning (PbCRL) tackles the issue of deriving safety constraints in reinforcement learning based on human preferences. Conventional Bradley-Terry models struggle to account for the asymmetric and heavy-tailed characteristics of safety costs, resulting in an underestimation of risk. PbCRL presents a more efficient solution that avoids limiting assumptions and the need for extensive expert demonstrations, enhancing its relevance to practical applications. This research, detailed in arXiv (2603.23565), emphasizes the cost-effective and dependable learning of intricate, subjective, and difficult-to-define safety constraints.