Dual-Agent DRL Framework for Robust Facility Coverage Under Interdiction

A new paper on arXiv proposes a Dual-Agent Deep Reinforcement Learning (DADRL) framework to solve the Maximal Covering Location-Interdiction Problem (MCLIP), a classic bi-level optimization problem in resilient infrastructure planning. The MCLIP involves an upper level that selects facility locations to maximize coverage, and a lower level that executes worst-case interdiction to minimize that coverage. The strong coupling and high combinatorial complexity of both levels make traditional methods ineffective. The DADRL framework uses adversarial learning with two agents: a location agent (upper level) and an interdiction agent (lower level). The location agent is trained against an evolving interdiction agent, allowing it to capture dynamic competitive interplay. The paper claims three contributions, including this adversarial training approach. The work is relevant to fields like emergency services, supply chains, and telecommunications where facilities must remain operational under attack.