Resilient Strategies for Stochastic Systems: How Much Does It Take to Break a Winning Strategy?

A team of researchers from institutions including the University of Texas at Austin and Technical University of Munich has published a foundational paper titled 'Resilient Strategies for Stochastic Systems: How Much Does It Take to Break a Winning Strategy?' The work, accepted to the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026), formally introduces the concept of resilience into the stochastic setting. It addresses a critical gap in AI and game theory: how to design decision-making strategies for agents (like robots or autonomous systems) that remain robust when their intended actions cannot be executed due to real-world disturbances, such as actuator failures or environmental interference.

The research provides a comprehensive set of fundamental problems centered on Markov Decision Processes (MDPs) and stochastic games with reachability and safety objectives. To quantify resilience in a probabilistic world, the authors propose various methods for aggregating the 'amount' of disturbance, such as measuring it in expectation or for the worst-case scenario. A key innovation is using quantitative measures, like the frequency of occurrence, to reason about potentially infinite sequences of disturbances. This framework allows developers to mathematically answer the titular question—how much disturbance is needed to break a strategy—enabling the design of AI systems that are provably robust against specified levels of failure.