Attention and Reward Hybrid Enhanced Reinforcement Learning for Multi-Agent Tasks
Peng Ju, Yan Kong · Unmanned Systems · 2025
Although traditional reinforcement learning algorithms have achieved notable success in multi-agent scenarios, persistent challenges including slow learning convergence and insufficient coordination capabilities persist in complex tasks. To overcome these limitations, this study proposes an optimized multi-agent reinforcement learning that integrates a multi-reward mechanism and a multi-head attention mechanism to improve collaborative efficiency and task completion rates in multi-agent environments. We innovate within the multi-agent deep deterministic policy gradient framework by introducing a multi-reward system with dynamic weight adjustments and a multi-head attention mechanism. The multi-reward system incorporates target occupation rewards, collaborative behavior rewards, and action smoothness penalties, where reward weights are dynamically adapted during training to optimize agent coordination. Concurrently, the multi-head attention mechanism enhances environmental perception and decision-making by processing observational data across multiple parallel attention heads, enabling agents to capture critical spatial and interactive features. We evaluated the method’s performance on both cooperative and competitive tasks, demonstrating superior learning efficiency and task completion quality compared to conventional approaches, along with enhanced adaptability and stability. These achievements thereby provide a theoretical foundation for applications in multi-robot cooperative systems and intelligent traffic scheduling scenarios.