Network Reduction Using Optimization-Reinforcement-Learning Model based on Flow Performance

Abstract

Modern infrastructure systems, such as transportation networks, energy distribution, and water networks, are becoming increasingly complex due to their expanding scale and dynamic demand profiles. Traditional network simplification techniques often rely on the preservation of topological features, which is not able, which is not able to adapt to operational dynamics or fluctuating flows. This limitation is especially pronounced in critical infrastructures, where maintaining functional performance under disruption or budget constraints is vital. This thesis proposes a novel optimization reinforcement learning (RL)-based framework for network reduction that prioritizes flow efficiency over structural preservation. Using the Minimum Cost Flow Problem (MCFP) as a performance proxy, we introduce a two-stage method that integrates scenario-based optimization with a Vanilla Policy Gradient (VPG) algorithm. The model is trained to generate sparse, efficient network topologies capable of replicating the behavior of a full-scale reference network under diverse demand conditions. The proposed method is validated on three real-world infrastructure systems: a natural gas network from Shelby County, Tennessee; a municipal water distribution system; and a synthetic power grid representative of regional electricity transmission. In all domains, the RL-designed networks consistently approximate the flow behavior of the original systems, even under strict arc constraints. Compared to the widely used Backbone heuristic - which requires approximately 85-90% of the original arcs to preserve performance, the RL model achieves convergence with only 60-70% of the network structure. Notably, the RL-based approach demonstrates earlier and smoother convergence, better alignment with cost distributions, and robustness across stochastic demand scenarios. In contrast, the heuristic Backbone method often overestimates flow costs under tight budgets and fails to reproduce distributional behaviors without full structural retention. In conclusion, this work demonstrates that optimization-guided reinforcement learning can serve as a powerful paradigm for reducing the complexity of infrastructure networks while preserving essential operational characteristics. The results highlight the promise of data-driven, adaptive reduction strategies as foundational tools for resilient and efficient infrastructure planning across energy, water, and gas domains.