Deep Reinforcement Learning for Autonomous Multi-Agent Systems in Real-Time Strategy Games

Abstract: Deep Reinforcement Learning (DRL) has emerged as a transformative approach for training autonomous agents, especially in complex, high-dimensional environments like real-time strategy (RTS) games. This study explores the use of DRL for autonomous multi-agent systems operating in RTS environments such as StarCraft II and Dota 2. We employ multi-agent reinforcement learning (MARL) architectures, including QMIX, MAPPO, and MADDPG, and compare their performance across coordination, task success, and adaptability metrics. Statistical analyses, including regression and predictive modeling, are used to determine how agent communication, environment stochasticity, and resource availability impact learning efficiency. Experimental results show that agents using centralized training with decentralized execution (CTDE) significantly outperform fully independent learners in both task completion and strategic flexibility. Keywords: Deep Reinforcement Learning, Multi-Agent Systems, Real-Time Strategy Games, QMIX, MAPPO, MADDPG, CTDE, PyTorch, TensorFlow, SHAP, LIME