Shields to Guarantee Probabilistic Safety in MDPs

A new academic paper submitted to arXiv introduces a formal framework for ensuring probabilistic safety in autonomous agents operating within Markov Decision Processes (MDPs). While classical shielding techniques guarantee that unsafe events never occur, this research addresses the more complex challenge of probabilistic safety, where undesirable outcomes are permitted within acceptable probability limits. The authors demonstrate that preserving strong guarantees on both safety and maximal permissiveness is impossible in this context. Consequently, they propose natural shields with weaker guarantees and introduce novel offline and online shield constructions that ensure strong safety standards. Empirical evaluations included in the study highlight the practical advantages and computational feasibility of these new shielding methods. This work represents a significant advancement in the field of AI safety, offering robust solutions for managing risk in autonomous systems where absolute safety cannot be guaranteed. The findings provide a conservative extension of classical shielding theories, aiming to balance safety requirements with operational flexibility in probabilistic environments.