TRIDENT Ensures Provably Safe Multi-Agent Reinforcement Learning

Safe coordination in complex networked cyber-physical systems presents a significant challenge for learning algorithms. These systems require simultaneous handling of hybrid discrete-continuous actions, strict training-time safety constraints, and adherence to physics-governed dynamics. The authors identify that these three elements create a "directed cycle of biases" that can undermine standard compositions of existing learning modules. To address this, the paper introduces TRIDENT, a novel Multi-Agent Reinforcement Learning (MARL) framework. TRIDENT's three core components are specifically co-designed to counteract these biases. It incorporates a Richardson-Romberg gradient correction to reduce Gumbel-Softmax bias, a Lyapunov-constrained sequential trust-region update to ensure feasibility at each iteration, and a physics-informed residual critic that decomposes value rather than reward. TRIDENT offers provable guarantees, including an O~(1/sqrt(K)) convergence rate to a constrained Nash equilibrium and an O(sqrt(K)) cumulative-violation bound. Evaluated on multi-UAV mobile-edge computing, autonomous intersection management, and a hybrid SMAC variant, TRIDENT dramatically cuts training-time violations by 95.5% compared to MADDPG and 76.3% over MACPO, while also improving rewards by 13.5% over the strongest unconstrained baseline.