Game-Theoretic Multi-Agent Reinforcement Learning for Economic Resource Allocation Optimization

Abstract

This paper presents a novel framework for optimizing economic resource allocation by integrating computational game theory with multi-agent reinforcement learning (MARL), addressing the challenges of dynamic, multi-agent interactions in complex economic systems. The framework leverages game-theoretic equilibrium concepts, such as Nash Equilibrium, alongside policy gradient methods and best-response dynamics to enable scalable, efficient, and stable decision-making in high-dimensional environments. An end-to-end experimental pipeline, validated using real-world data from the World Bank Open Data repository, demonstrates the effectiveness of the proposed approach. Quantitative results show that the framework achieves an economic utility score of 92.5,(±3.2), outperforming baseline models including Single-Agent RL (78.3), Non-Cooperative Game Theory (85.1), and Centralized Optimization (88.7). It also reduces convergence time to 750,(±25) steps and improves fairness, with a Gini coefficient of 0.15,(±0.02), indicating balanced resource distribution. Compared to existing models, the proposed method delivers superior policy stability (0.01 ± 0.005) and faster adaptation. These results highlight the framework’s ability to discover equitable, high-utility resource allocations while maintaining long-term equilibrium, making it a powerful tool for applications in market competition, supply chain management, and public policy optimization.