Contract-Based Compositional Shielding for Safe Multi-Agent Reinforcement Learning

In three linesDecentralised shielding method for multi-agent reinforcement learning ensuring global safety without centralised runtime control. Agents share a global LTL_safe specification and select local obligations whose conjunction implies the global specification, via a non-stationary multi-armed bandit. Evaluation across 6 environments and 15 algorithmic variants.

Summary generated by Claude — human-verified