Evolutionary Hypergraph Dismantling via Deep Reinforcement Learning

Abstract Assessing the vulnerability of complex systems requires effective hypergraph dismantling strategies, yet existing methods struggle with the dynamic nature of cascading failures and the rugged optimization landscapes of high-order networks. In this paper, we propose a novel framework, Evolutionary Hypergraph Dismantling via Deep Reinforcement Learning (HD-EDR). First, we model a realistic dismantling environment incorporating hyperdegree-based and residual-capacity-based load redistribution mechanisms. Second, we introduce a hybrid learning architecture that synergizes the global exploration of Evolutionary Strategies with the gradient-based exploitation of DRL. A bidirectional parameter synchronization mechanism is designed to prevent the agent from being trapped in local optima. Furthermore, we integrate an inductive encoder to capture the evolving high-order dependencies of the residual network in real-time. Extensive experiments across nine real-world datasets demonstrate that our framework significantly outperforms stateof-the-art baselines, providing a highly effective and robust strategy for maximizing structural damage in high-order networks.