Learn to decompose multiobjective optimization models for large‐scale networks

Abstract

AbstractInfrastructures can be modeled as large‐scale networks consisting of nodes and arcs, making network optimization a popular modeling option for arising problems. In specific, providing timely restoration plans for interdependent infrastructures facing disruptions has been a challenge for decision makers. In this study, we focus on geospatial (co‐location) and functional interdependencies to capture the impact of cascading failures on infrastructure systems. The dynamics of real networks are more complicated to be captured by one objective function. Therefore, we define three objective functions in three pillars of sustainability: (a) economic, (b) social, and (c) environmental. To solve the multiobjective optimization model, we develop a learn‐to‐decompose framework, consisting of a multiobjective evolutionary algorithm based on decomposition module and a Gaussian process regression (GPR) module to periodically learn from the obtained Pareto front and guide the search direction. We also included a heuristic module to address two significant challenges in restoring interdependent infrastructures: the island scenario and co‐location interdependencies. We applied the proposed framework to benchmark problems and interdependent water and transportation networks in the City of Tampa, FL. We carried out sensitivity analyses to monitor the performance of the GPR by different kernel functions. We also provided insights for decision makers by finding the trade‐off between fortification (proactive) and restoration (reactive) costs. The result demonstrates the proposed framework is feasible and applicable for large‐scale networks.