Knowledge Transfer for Multiagent Reinforcement Learning Based on Shared Advice Archive

Experience sharing has developed as a popular knowledge transfer method for multiagent reinforcement learning, which is able to enhance the learning speed. To further enhance the data efficiency and reduce the involved communication, this paper proposes a novel knowledge transfer framework for multiagent reinforcement learning based on a shared advice archive, termed KT-SAA. By decoupling shared advice from individual agents, KT-SAA enhances data efficiency and enables effective cross-task knowledge transfer. A new trigger condition is also proposed to balance exploration and exploitation, significantly reducing communication overhead in the early learning stages without compromising performance. Extensive simulations demonstrate that KT-SAA achieves competitive or superior learning performance with less communication than competitive methods, indicating the effectiveness and efficiency of the proposed method.