Deep Reinforcement Learning-Based Equitable Post Disaster Resource Allocation Incorporating Social Vulnerability

Author

• Trevor John Mogaka, Southern Illinois University CarbondaleFollow

Date of Award

5-1-2026

Degree Name

Master of Science

Department

Civil Engineering

First Advisor

Sen, Debarshi

Abstract

Prior Deep Reinforcement Learning (DRL) approaches frame post‑disaster recovery as a sequential decision-making problem over infrastructure elements, using graph-based system models and resilience-oriented rewards. This thesis builds on an existing DRL framework by introducing a modified CDC Social Vulnerability Index (mCDC-SVI), developed using Max-min scaling, to quantify social vulnerability.A synthetic interdependent infrastructure testbed is divided into four socio‑economic regions using federal poverty and SNAP eligibility guidelines. The DRL simulation and Deep Q‑Network architecture are retrained under a multi-objective reward function that is efficiency-equity weighted (α:β). Agents are evaluated across unconstrained and constrained budget scenarios to examine how varying α and β influence their repair prioritizations, and then compared with rules-based policies to examine the trade-offs arising between social vulnerability reduction and system functionality optimization. The results show that all agents attain an overall system functionality of 1.0 across all budgets. Under constrained budgets, equity policies (higher β) have the highest social vulnerability reduction rate in the most vulnerable region while also enabling faster recovery. Efficiency‑weighted policies (higher α) prioritize system functionality restoration at the cost of persistent disparities. In the 58-unit budget scenario, the fully-efficiency weighted DQN agent (αc: βc = 100%:0%) has the highest average functionality (0.79), exceeding the best rules-based agent (Strategic Agent). The DQN agents learn repair sequences that move vulnerable regions above the poverty line through discrete improvements in social vulnerability. Unlike the rules-based Strategic Agents, the DQN agents reveal equity-efficiency trade-offs: the αc: βc = 25%:75% configuration has an overall functionality gap of 0.05 relative to the αc: βc = 100%:0%, while achieving the steepest mCDC-SVI reduction in the most vulnerable region.