Intelligent Decision Making in Dynamic Environments Based on Evolutionary Game Theory and Multi-Agent Reinforcement Learning
Yuan Zhang, Jiangnan Wang, Xuan Li, Chao Li · 2025
As dynamic environments become more complex, the need for autonomous decision making in high-dimensional state and action spaces increases. In this paper, we propose a multi-intelligent deep reinforcement learning (MADRL) framework based on evolutionary game theory and designed for parallel training in multiple environments. The framework addresses the challenges of algorithmic overfitting, slow convergence, and inefficient training by integrating multi-reward systems and introducing selection and mutation mechanisms inspired by collective intelligence. These enhancements improve learning efficiency and exploration, while ensuring robustness against adaptive adversaries through self-play training. Simulation result show that the proposed framework offers greater adaptability, faster convergence, and lower hyperparameter sensitivity than the classical approach. The results show that this approach enables agents to develop effective strategies and autonomous decision-making capabilities. This highlights its potential for applications in complex and dynamic systems such as robotics, autonomous navigation, and multi-agent coordination.