Generalizable agent modeling for agent collaboration-competition adaptation with multi-retrieval and dynamic generation

Chenxu Wang, Yonggang Jin, Cheng Hu, Youpeng Zhao, Zipeng Dai, Jian Zhao, Liuyu Xiang, Junge Zhang, Zhaofeng He

Abstract: Highlights•This research focuses on single-agent generalization in long-sequence tasks, aiming to enable dynamic adaptation to tasks and multi-strategy agent interactions.•A comprehensive framework with dynamic retrieval for policy network parameters is proposed to enhance agent generalization in diverse scenarios.•Experiments across SMAC, Overcooked-AI, and Melting Pot validate the proposed method’s superior performance over existing research.•The novel MRDG method integrates MAS agents’ actions/attributes into a knowledge base, boosting adaptability and generalization.•This work enriches AI theoretical frameworks and enables versatile single-agent applications across multiple fields.

External IDs:dblp:journals/ijon/WangJHZD0XZH25