TRIDENT Achieves Provably Safe Multi-Agent Reinforcement Learning

Ensuring safe coordination in complex cyber-physical systems, especially those involving multiple agents, presents significant challenges for reinforcement learning. These systems often combine discrete and continuous actions, demand strict safety constraints during training, and operate under physics-governed dynamics. Naive combinations of existing learning modules fail due to a "three-way coupling" of biases among these features. TRIDENT is the first Multi-Agent Reinforcement Learning (MARL) framework specifically engineered to break this coupling. It features three co-designed components: a Richardson-Romberg gradient correction to reduce Gumbel-Softmax bias, a Lyapunov-constrained sequential trust-region update to ensure per-iterate feasibility, and a physics-informed residual critic that decomposes value rather than reward. The framework provides provable guarantees, including an O~(1/sqrt(K)) convergence rate to a constrained Nash equilibrium and an O(sqrt(K)) cumulative-violation bound. Applied to scenarios like multi-UAV mobile-edge computing and autonomous intersection management, TRIDENT drastically cuts training-time safety violations (e.g., 95.5% over MADDPG) while simultaneously improving rewards.