Multi-agent deep reinforcement learning for resilience optimization of building structures considering utility interactions for functionality

The resilience optimization of the built environment under extreme events with building structures and interdependent physical infrastructure systems is inefficient due to the large action spaces of pre-hazard mitigation alternatives. Hence, a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) based resilience optimization framework is proposed herein to enhance the resilience of the built environment with large mitigation action spaces. The proposed framework is divided into two parts including (1) Performance-Based Environment (PBE) to assess the resilience of the building structures under given hazard scenarios considering dependencies and interdependencies of the physical infrastructure systems, and (2) MADDPG for resilience optimization of the interdependent building structures given minimal mitigation and repair costs. The PBE is developed to assess the consequences of the mitigation actions under hazard by utilizing steps including component consequence and recovery assessment, network modeling and functionality assessment, and portfolio consequence and resilience assessment. The multi-agents for resilience optimization utilize Deep Deterministic Policy Gradients (DDPG) including actor and critic networks to learn to optimize the mitigation actions by maximizing the cumulative reward function given the hazard scenarios. The proposed framework is illustrated on a built environment with building structures and two physical infrastructure systems including a Water Network (WN) and an Electrical Power Network (EPN) system. The provided framework efficiently optimizes complex, interdependent, and large infrastructure systems under extreme events by utilizing physics-based environments and artificial intelligence.