Deep Reinforcement Learning for Network Service Recovery in Large-Scale Failures

In large-scale failures, telecommunication carriers are required to recover network services as quickly as possible. Since few resources are available in the early stage of the disaster, it is desirable to be able to recover services by repairing as few communication facilities as necessary. This is a combinatorial optimization problem. Recently, methods for solving the network problem using reinforcement learning have been studied. However, existing methods are not sufficient to recover network services because they use only network information for agent learning and decision making. We propose a deep reinforcement learning method that utilizes the services information as well as the network information for minimizing the total repair cost. We represent communication networks and network services as a supply graph and a demand graph, respectively, and use them for agent learning and decision making. Numerical experiments show that the proposed method can reduce the total cost from baselines.