Reinforcement Learning in the Maintenance of Civil Infrastructures

Abstract

Life-cycle management of aged civil infrastructures is an issue of worldwide concern. The process of sequential decision making on structural maintenance is usually considered as a Markov Decision Process (MDP) where Markov property holds in the structural condition transition due to the deterioration and maintenance. However, policy-making for large MDPs for maintenance of complex realistic infrastructures has long been a challenging problem due to the high-dimensions. Thus we introduce a deep reinforcement learning(DRL) framework to make this available, and a deep Q network implemented by CNN is employed to approximate the state-action value in the high-dimensional state-action space. A maintenance task of a cable-stayed bridge is designed and used to verify the efficiency of the proposed approach. The results show that the DRL is effective and efficient at the policy-making for maintenance tasks of complex civil infrastructures with high-dimensional state-action space.