Multi-agent reinforcement learning and variational inequality models for international trade networks under crisis

Global disasters increasingly disrupt agricultural commodity flows, with food insecurity as a major consequence. Quantitative tools to assess such impacts are essential for resilience. We propose a hybrid Multi-Agent Reinforcement Learning (MARL) architecture to solve Variational Inequality (VI) problems in multi-commodity trade equilibria. While variational inequalities offer a rigorous method, their resolution via MARL faces stability and convergence challenges. Our actor–critic approach integrates a Gradient-based Learning Rate (GLR) scheduler, adaptive epsilon decay, prioritized replay, and a dual reward combining individual and centralized feedback. Agents representing supply and demand learn optimal strategies to reach equilibrium in simulated markets. Experiments, spanning stable conditions, dynamic price shifts, and route blockages, show faster convergence and stronger robustness than Multi-Agent Proximal Policy Optimization (MAPPO) and Multi-Agent Deep Deterministic Policy Gradient (MADDPG). Results highlight the promise of MARL for simulating economic behavior and optimizing decentralized decision-making in complex systems.