Agent Decision-Making Through Human Visual Attention Prediction and Deep Reinforcement Learning in a Competitive Wargame: An Experimental Study

In competitive wargames, human players leverage visual attention mechanisms to rapidly identify mission-critical information and make real-time decisions. However, conventional deep reinforcement learning (DRL) methods often lack this ability to focus on critical cues, limiting their efficiency and decision quality. To address this gap, we propose a novel framework that enhances agent decision-making by incorporating human visual attention (HVA) into DRL through two key innovations. First, we develop an improved convolutional neural network for visual attention prediction, which outperforms existing models across multiple evaluation metrics. Second, we propose HVA proximal policy optimization (PPO), a novel DRL framework that integrates HVA through two mechanisms: 1) attention-guided sampling; and 2) multimodal feature fusion for comprehensive state representation. Experimental results demonstrate that our framework achieves an 84.4% winning rate (15.6% higher than standard PPO) with significantly faster convergence. Ablation studies further confirm that both the attention prediction model and its integration mechanism contribute substantially to performance gains. These findings highlight the effectiveness of incorporating human cognitive patterns into agent learning, offering new insights for human-agent hybrid strategies in dynamic tasks.