AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-22
Reinforcement Learning for Global-Local Control of Dynamic Networks with Applications to Systemic Risk
The authors propose a reinforcement learning framework for control problems on dynamic networks, where decisions combine graph-level global actions with node-level local actions. A message-passing neural network with separate global and local policy heads yields policies satisfying permutation symmetries and transparency constraints, and explicit Lipschitz bounds provide convergence guarantees for a projected policy-gradient algorithm. Applied to systemic risk management in financial networks, where a regulator allocates capital interventions across institutions, the architecture outperforms standard feedforward baselines and scales to large networks.