Multi-Agent Collaboration for Autonomous Systems Based on Implicit Strategy Prediction and Prioritization

Collaborative multi-agent systems (MAS) have the potential to greatly enhance task performance by enabling agents to learn and adapt through interaction. However, traditional methods often face inefficiencies, particularly when agents encounter similar tasks repeatedly. To address these challenges, we propose a novel multi-agent collaboration scheme (MACS) based on deep reinforcement learning (DRL), aimed at improving interaction efficiency and reducing redundant learning. The proposed framework integrates two main components: the Implicit Strategy Prediction Algorithm (ISPA) and the Optimized Experience Prioritization Algorithm (OEPA). ISPA focuses on predicting agents' underlying strategies through their actions, ensuring that experience learning remains both unique and relevant. OEPA then prioritizes strategies based on their effectiveness, guiding agents to rapidly adapt and learn in downstream tasks by leveraging optimized strategies from long-term exploration. The framework's efficacy is demonstrated through simulations across multiple environments, where it achieves faster task completion and enhanced learning efficiency compared to traditional approaches.