New Shielding Method Enhances Multi-Agent AI Safety

In multi-agent reinforcement learning (MARL), ensuring safe coordination is a significant challenge, especially when global safety cannot be enforced by a single agent. Traditional decentralized shields often restrict optimal team behaviors that are only safe through coordinated actions. This new research proposes a contract-based compositional shielding method to address this, aiming to recover team-optimal safe behavior while maintaining decentralized execution. The core of the approach involves agents sharing a global safety specification, expressed in Linear Temporal Logic (LTLsafe). Instead of individual agents enforcing global safety, they select tuples of local LTLsafe obligations. The conjunction of these local obligations is designed to imply the global specification. Each agent can then rely on the other agents' local obligations as assumptions, as the entire contract tuple is certified simultaneously, allowing for projection into local action masks. During the learning phase, a non-stationary multi-armed bandit algorithm is used to choose among a library of local LTLsafe obligations. This selection process aims to optimize team reward without compromising end-to-end safety. The method was evaluated across six environments and fifteen algorithmic variants, demonstrating its effectiveness in achieving safe, coordinated behavior in complex multi-agent systems.