Deep Reinforcement Learning for Life-cycle Maintenance Planning of Large Infrastructure Networks

This paper proposes a Deep Reinforcement Learning (DRL) framework designed to derive an optimal sequence of maintenance actions for large infrastructure networks with interacting assets. The DRL agent learns an adaptive policy that maps the system's belief state, i.e., a probability distribution over the true network condition, to optimal maintenance actions. The agent is trained using Proximal Policy Optimization (PPO) with a centralized belief and action architecture. Furthermore, the concept of curriculum learning is employed, where the agent is progressively trained on environments of increasing complexity, transferring the acquired knowledge of the simpler configurations to the subsequent harder ones. The framework is tested on Amsterdam's quay wall network, demonstrating significantly lower expected life-cycle costs compared to heuristic-based plans.

Keywords: Deep Reinforcement Learning, Infrastructure Maintenance, Proximal Policy Optimization, Curriculum Learning, Multi-component Deteriorating Systems.