TRIDENT — safe multi-agent RL breaks hybrid-dynamics coupling
arXiv paper formalizes why naive composition of discrete-action, safety-constrained, and physics-governed modules fails in networked cyber-physical systems — a three-way feedback loop that standard algorithms can't escape.
TRIDENT co-designs three components to cancel each leak: Richardson-Romberg gradient correction (reduces Gumbel-Softmax bias quaradratically), Lyapunov-constrained trust-region updates (per-iterate feasibility), and physics-informed residual critic (decomposes value estimation).