Formal Verification of Learned Multi-Agent Communication Policies via Decision Tree Distillation

The article presents a novel framework for the formal verification of learned multi-agent communication policies by distilling neural policies into decision trees, achieving a fidelity of 97.9%. The proposed four-stage pipeline includes feature extraction, decision tree distillation, translation to PRISM specifications, and verification of Probabilistic Computation Tree Logic properties, demonstrating an 88.9% satisfaction rate for 18 verified temporal logic properties in multi-drone coordination scenarios. This framework offers a practical solution for ensuring safety in safety-critical applications of multi-agent reinforcement learning, effectively bridging the gap between deep learning and formal verification.